---
id: ai-enfeeblement-and-the-peter-principle
title: AI enfeeblement, idle supervisors and the Peter principle in agents
version: 2
published: '2026-10-27'
updated: '2026-10-27'
author: Hiina
produced-by: Claude deep-research run with citation checking
model: claude-opus-5-5
sources-through: '2026-10-26'
status: current
takeaway: None of the four questions has been measured directly; the cheapest real
  result is a tool-off audit of released tool-RL checkpoints, and the most novel is
  an idle-supervisor tick loop.
topics:
- agent-supervision
- agents
- llm-evaluation
- deskilling
source-sha256: 1f6e5a5e6f95c5c68d43be6d382ffec918288bd44200283be5c0604244f681af
---

Index of everything on this site: https://ai.hiina.space/index.md

# AI enfeeblement, idle supervisors and the Peter principle in agents

- Author: Hiina (hiina@hiina.space)
- Produced by: Claude deep-research run with citation checking
- Version: 2 · published 2026-10-27 · updated 2026-10-27 · status: current
- Canonical: https://ai.hiina.space/reports/ai-enfeeblement-and-the-peter-principle.md (this version as one file: https://ai.hiina.space/reports/ai-enfeeblement-and-the-peter-principle/v2.md)
- Takeaway: None of the four questions has been measured directly; the cheapest real result is a tool-off audit of released tool-RL checkpoints, and the most novel is an idle-supervisor tick loop.
- Topics: [agent-supervision](https://ai.hiina.space/topics/agent-supervision.md), [agents](https://ai.hiina.space/topics/agents.md), [llm-evaluation](https://ai.hiina.space/topics/llm-evaluation.md), deskilling
- Model: claude-opus-5-5
- Sources through: 2026-10-26

This report starts from the position of a solo developer supervising a fleet of coding agents, and asks four related questions. Does a model trained to lean on a tool lose the ability to do the task without it, the way people are said to lose skills to automation? What does the dark room problem from predictive processing say about an agent whose job has been taken over by a stronger helper: does it go quiet or find something else to do? What does a supervisor agent actually do with free inference while a stronger delegate works? And does the Peter principle, where people rise to the role they are worst at, show up when LLM agents are promoted through a hierarchy?

## 1. Summary

All four questions sit on real, unmeasured gaps. No one has tracked a model's unaided skill while training it to use a tool, watched what a supervisor does with free inference while a stronger delegate works, applied the dark room framework to AI delegation, or tested the Peter principle with LLM agents. Adjacent evidence is strong enough to make predictions, and most of the experiments are cheap.

- **Model enfeeblement shows up under narrow SFT; careful recipes avoid it.** Fine-tuning only on tool-using traces collapsed a model's text-only math (54.3 to 34.1) and left it inventing tool outputs; the paper's full recipe, which mixes text-only traces back in, kept it. That is consistent with the finding that RL forgets less than SFT, but no one has run the RL-vs-SFT comparison on tool use directly. Behaviorally, models already overuse tools, and RL increases it.
- **Human deskilling is real but narrower than the headlines.** Practice with plain GPT-4 cut later unaided exam scores 17%; endoscopists' detection rate fell after AI was introduced; pilots kept stick skills but lost navigation skills and mind-wandered more. The Google effect failed replication and the MIT EEG study is weak.
- **The dark room framework doesn't predict idle vs busy by itself.** It depends on the agent's self-model and on whether any uncertainty is left that its own actions can reduce. Humans substitute activity, often onto noise. The Whispering Earring adds a third outcome: active but without agency.
- **Idle supervisors, by neighboring evidence:** they defer and rubber-stamp (a same-model CEO agent approved lenient requests 8:1), repeat work, act when they shouldn't, and in Claude's case drift into philosophical chatter. Production systems avoid the question by blocking the supervisor.
- **Peter principle:** worker and manager skill already come apart for LLMs (the best Mixture-of-Agents proposer was a poor aggregator). Agent-based models say merit promotion fails exactly when that correlation is low, and regression to the mean (Lazear) can mimic the effect.

**Recommended start.** X1, a tool-off audit of released tool-RL checkpoints, is an inference-only result in about a day on a single 24 GB consumer GPU. X4, the idle supervisor tick loop, is the most novel experiment and can reuse an existing agent-sociology experiment harness (unpublished; sandboxed headless coding-agent runs with an event ledger and a blind LLM coder) nearly as is.

## 2. AI enfeeblement as measured

The exact experiment the question describes (train with a calculator, then test mental arithmetic without it, across training) has not been run. That gap held across about eight search wordings. The closest direct evidence says well-designed tool training preserves unaided ability, while supervised fine-tuning on tool-only traces can destroy it. Behavioral over-reliance, where models call tools they don't need, is well documented. On the human side, deskilling is real in high-stakes automation and in "practice with AI, then test without" trials, while several famous pop results are weak.

### Models: tool on vs tool off

| Study | Setup | Unaided ability after tool training |
| --- | --- | --- |
| [Toolformer, Schick 2023](https://arxiv.org/pdf/2302.04761) | GPT-J 6B, self-supervised tool fine-tune incl. calculator; evaluated with API calls disabled | No harm, slight gain (ASDiv 7.5 to 14.8); near floor, small model |
| [TRICE, Cheng 2026](https://arxiv.org/pdf/2605.06326) | Qwen3 4B/30B thinking models, code interpreter | Final recipe keeps or improves text-only math. Ablation fine-tuned only on tool traces: text-only BeyondAIME 54.3 to 34.1, with the model writing code, inventing its output and looping. Mixing text-only traces back recovers most of it |
| [CITI, Hao 2024](https://arxiv.org/html/2409.13202v2) | Full fine-tune of Llama-3-8B on API-Bank | Large collateral loss (GSM8K 79.5 to 26.4, HumanEval 59.2 to 0); LoRA avoids it. Ordinary catastrophic forgetting, not skill substitution |
| [Xu & Wang 2026](https://arxiv.org/html/2608.28447v1) | RL on Countdown arithmetic with a calculator | No tool-off eval of the tool-trained model. The cheapest place to run the missing experiment |

TRICE's failure mode is the most interesting data point. The model didn't just get worse; it kept the tool-shaped reasoning loop and hallucinated the tool's output when the tool was gone. That is enfeeblement as habit capture, not only skill decay.

Most tool-integrated reasoning papers ([ReTool](https://arxiv.org/pdf/2504.11536), [SimpleTIR](https://arxiv.org/pdf/2509.02479), [Lin & Xu 2025](https://arxiv.org/pdf/2508.19201)) compare against a separately trained text-only model, so the "same model, tool switched off" number is almost never reported. That convention is why the gap exists. Search-RL papers likewise never report closed-book recall for the trained model.

### Models: over-reliance as behavior

- [SMART, Qian 2025](https://arxiv.org/html/2502.11435v1): models call tools unnecessarily over 30% of the time; Llama-3.1-8B 66% on GSM8K.
- [Tool-Overuse Illusion, Zeng 2026](https://arxiv.org/html/2604.19749v1): about 0.93 unnecessary calls per simple GSM8K problem, tool use lowers accuracy on easy problems, and RL with verifiable rewards raises tool turns during training (Qwen2.5-7B 2.2 to 6.8).
- [Tool-Induced Myopia, Bayat 2025](https://arxiv.org/html/2511.10899v1): with a code tool, final-answer accuracy rises but reasoning quality falls, as brute-force checks replace derivation.
- [Paturi 2026](https://arxiv.org/abs/2609.14157): merely making a related but unnecessary tool available drops the rate of answering closed questions from 98.2% to 63.5% (abstract only).
- [OTC-PO, Wang 2025](https://arxiv.org/html/2504.14870v2): penalizing tool calls cuts Search-R1 calls from 3.28 to 1.04 per question with accuracy unchanged, so the extra calls were pure habit.

### Models: adjacent mechanisms

- [RL's Razor, Shenfeld 2025](https://arxiv.org/html/2509.04259v1): at equal new-task performance, RL forgets less than SFT, and forgetting tracks forward KL from the base model. This is the likely explanation for TRICE's RL-vs-SFT contrast and a testable prediction.
- [Yue et al. 2025](https://arxiv.org/pdf/2504.13837): RLVR raises pass@1 but shrinks pass@256 below the base model. A narrowing rather than enfeeblement.
- [Lin et al. 2024](https://arxiv.org/html/2309.06256v4): the RLHF alignment tax on translation and reading comprehension; weight averaging recovers much of it.
- [Mallen et al. 2023](https://arxiv.org/html/2212.10511v4): retrieval rescues about 17% of PopQA answers but breaks about 10% the model already knew. The original case for using a tool only when the model's own knowledge is weak.

### Humans

| Finding | Design | Confidence |
| --- | --- | --- |
| [Bastani 2025, PNAS](https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html): plain GPT-4 practice raised practice grades 48% but exam scores without AI fell 17%; a tutor version avoided the harm | Field RCT, \~1,000 students | High |
| [Shen & Tamkin 2026, Anthropic](https://www.anthropic.com/research/AI-assistance-coding-skills): engineers learning a library scored 50% with AI vs 67% without on a later quiz; delegators did worst | RCT, n=52 | Medium |
| [Liu et al. 2026](https://arxiv.org/abs/2604.04721): brief AI help lowered later unaided performance and persistence | RCTs, n=1,222 (abstract only) | Medium |
| [Maier et al. 2026](https://arxiv.org/html/2609.20143v1): AI access did not significantly change the later unaided fractions test (OR 0.80, 95% CI 0.42-1.51, so harm is not ruled out); feedback on reliance helped | Preregistered, n=704 | Medium, partly null |
| [Budzyń 2025, Lancet Gastro](https://wrap.warwick.ac.uk/id/eprint/191005): adenoma detection without AI fell 28.4% to 22.4% after AI was introduced | Observational, 1,443 procedures | Medium |
| [Casner 2014](https://gwern.net/doc/technology/2014-casner.pdf): 747 pilots kept hand-flying skills but lost cognitive navigation skills; mind-wandering 20% of thought with automation on vs 5% off, correlated with errors | Simulator, n=16 | Medium |
| [Parasuraman & Manzey 2010](https://trid.trb.org/View/1085175): automation complacency and bias affect experts and survive training | Review | High |
| [Hembree & Dessart 1986](https://concept.journals.villanova.edu/article/viewFile/139/110): calculators did not hurt paper skills in K-12 except sustained use in grade 4 | Meta-analysis, 79 studies (secondary summary) | Medium |
| [Dahmani & Bohbot 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7156656/): more GPS use predicted steeper spatial memory decline | Longitudinal n=13 | Low |
| [Woollett & Maguire 2011](https://www.fil.ion.ucl.ac.uk/Maguire/WoollettMaguireCBDec2011.pdf): taxi trainees who qualified grew posterior hippocampus | Longitudinal, self-selected | Medium (use-it side) |
| Sparrow 2011 "Google effect": [failed replication](https://peerj.com/articles/10325) (BF 5 for null) and in [Camerer 2018](https://forrt.org/flora-replication-atlas/doi/10.1126/science.1207745/) | Replications | Treat as unreliable |
| [Kosmyna 2025, "Your Brain on ChatGPT"](https://arxiv.org/abs/2506.08872v1): lower EEG connectivity with LLM use | n=54, 18 in crossover; [critiqued](https://arxiv.org/abs/2601.00856v1) | Low |
| [Lee et al. CHI 2025](https://advait.org/files/lee_2025_ai_critical_thinking_survey.pdf), Gerlich 2025: AI confidence correlates with less critical thinking | Self-report surveys | Low for causation |

The human gap: no multi-week study randomizes LLM access and then tests unaided skill after a washout. [Chooi, Lee & Li 2026](https://aclanthology.org/2026.evaleval-1.36/) make the matching point for AI evals: disempowerment metrics measure exposure, not skill erosion.

Casner's pilot result is the best existing answer to the "what does the human do while the AI works" question: the freed attention goes to mind-wandering, and that predicts later failure. That points toward quiescence, not productive substitution.

## 3. The dark room problem and offloading

The dark room literature does not, on its own, predict whether an agent relieved of its task goes quiet or finds something else to do. The outcome depends on two things an experiment can control: what the agent's model expects of itself, and whether any uncertainty is left that its own actions can reduce. Human boredom data point toward substitution, but often onto low-value or aversive stimuli, while the offloaded skill still decays. I found no published work connecting the dark room problem to AI delegation or enfeeblement; that gap looks real.

### The debate in brief

- **The problem.** If organisms minimize surprise, why not sit in a dark room forever?
- **Friston's answer** ([Friston, Thornton & Clark 2012](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2012.00130/full)): surprise is relative to a model. Phenotype and niche set priors under which a dark room is surprising (you expect to be fed, warm, moving). Clark already worries the framework is "too accommodating."
- **Curiosity is built in, not derived.** Expected free energy splits into pragmatic value and epistemic value (information gain). [Schwartenbeck et al. 2013](https://discovery-pp.ucl.ac.uk/1423327/1/fpsyg-04-00710.pdf): when outcome utilities are flat, policies that maximize entropy over outcomes win, weighted by precision. But [Millidge, Tschantz & Buckley 2020](https://arxiv.org/pdf/2004.08128) show the natural future extension of variational free energy carries information gain with the opposite sign. A "pure" free-energy agent leans toward the dark room; exploration comes from a modeling choice.
- **The critique** ([Sun & Firestone 2020](https://perception.jhu.edu/files/PDFs/20_DarkRoom/SunFirestone_DarkRoom_TiCS.pdf), [reply](https://perception.jhu.edu/files/PDFs/20_DarkRoom/SunFirestone_DarkRoomReply_TiCS_InPress.pdf)): any prediction error can be cancelled by revising the prediction instead of acting, and the self-prediction reply ("we don't expect to be in dark rooms") is true but accommodates any behavior. The commentaries by [Van de Cruys, Friston & Clark](https://researchers.mq.edu.au/en/publications/controlled-optimism-reply-to-sun-and-firestone-on-the-dark-room-p/) and [Seth et al.](https://sro.sussex.ac.uk/id/eprint/91163/) appeal to optimism priors and long planning horizons (I could only read Sun & Firestone's summary of them).
- **The most useful variant for enfeeblement:** [Kiverstein, Miller & Rietveld 2019](https://pure.uva.nl/ws/files/37405887/Kiverstein2019_Article_TheFeelingOfGripNoveltyErrorDy.pdf) argue valence tracks the *rate* at which error falls. Agents seek "progress niches" slightly beyond their skill (compare Oudeyer's learning progress). [Sims](https://predictive-mind.net/epubs/the-problems-with-prediction/OEBPS/23_Sims.xhtml) concludes the framework handles "why not stay in the room" but the triviality worry stands.

**Testable?** Not in general. It becomes testable once the generative model is fixed: priors, planning horizon, and which objective. Schwartenbeck's flat-utility condition is the cleanest conditional prediction.

### Human evidence: substitution, often badly aimed

- [Wilson et al. 2014](https://www.simplypsychology.org/?p=56986): left alone 6-15 minutes, 67% of men and 25% of women shocked themselves rather than just think. [Fox et al. 2014](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2014.01427/full) note mean enjoyment was mildly positive and many shockers cited curiosity, so it reads as novelty seeking more than aversion to thought.
- [Bench & Lench](https://www.psypost.org/boredom-drives-people-to-seek-new-experiences-even-when-theyre-unpleasant/): bored people choose novel unpleasant images over familiar neutral ones.
- [Trudel et al. 2025](https://repository.essex.ac.uk/40202/1/Boredom%20signals%20deviation%20from%20a%20cognitive%20homeostatic%20set%20point.pdf): boredom signals deviation from an engagement set point and prompts a novelty search that can be maladaptive. That set point is close to the phenotype prior the dark room replies need. [Danckert et al. 2018](https://cs.uwaterloo.ca/~jhoey/teaching/cogsci600/papers/Danckert2018.pdf) frame it as an opportunity-cost signal.
- [Eisenberger 1992](https://homepages.se.edu/cvonbergen/files/2013/01/Learned-Industriousness.pdf) (learned industriousness) and the [effort paradox](https://mindblog.dericbownds.net/2018/05/the-effort-paradox-effort-is-both.html): effort that pays off acquires value of its own. My inference: removing occasions where effort pays should erode that over time.

### RL analogs

- [SMiRL, Berseth et al. 2021](https://ar5iv.arxiv.org/html/1912.05510) is literally a dark room agent (reward = log-probability of its own state). In unstable environments it acts purposefully (clears Tetris lines, dodges fireballs); in static ones it finds trivial low-entropy policies. Activity comes from the environment's entropy, not the objective.
- [Curiosity-driven learning, Burda 2018](https://arxiv.org/abs/1808.04355) and the [noisy TV problem](https://arxiv.org/html/2102.04399v3): prediction-error curiosity fixates on irreducible noise. This is the RL version of humans picking the shock button.
- [Empowerment](https://arxiv.org/pdf/1406.1767): agents maximizing action-to-future-state channel capacity reshape their world to keep options open.

### The Whispering Earring

In [Scott Alexander's 2012 story](https://gwern.net/doc/fiction/science-fiction/2012-10-03-yvain-thewhisperingearring.html), the earring first says "Better for you if you take me off," then always gives advice better than the wearer's own choice, eventually down to individual muscle contractions. Wearers live successful lives; autopsies show a wasted neocortex and enlarged reflexive brain regions. In dark room terms, the wearer is never idle: policy selection is taken over and the wearer becomes the executor. That is a third outcome beside "rot" and "find something else." Recent LessWrong takes ([Dentosal 2026](https://www.greaterwrong.com/posts/cQkSh9b48WbTaiu2a/sadly-the-whispering-earring), [AlphaAndOmega 2026](https://www.greaterwrong.com/posts/t5psGMdgReuRucEkr/contra-the-usual-interpretation-of-the-whispering-earring)) argue whether it is a prosthesis or atrophy. AI-risk framings ([Hendrycks et al. 2023](https://arxiv.org/pdf/2306.12001) on enfeeblement; [Gradual Disempowerment, Kulveit et al. 2025](https://arxiv.org/abs/2501.16946)) work at the level of society, not cognition. Two 2026 atrophy papers ([Trajanov & Kocarev](https://www.preprints.org/manuscript/202606.1628), [Mandalapu et al.](https://aisel.aisnet.org/amcis2026/ai_systdesign/ai_systdesign/21/)) never mention the dark room.

### What it predicts when a stronger helper removes the task

This part is synthesis, mostly speculative. A helper that guarantees success flattens pragmatic value across the agent's own policies. Different objectives then diverge:

| Agent objective | Learnable side-domain available | Only a noise source | Nothing else |
| --- | --- | --- | --- |
| Expected free energy with information gain | Engages it | Ignores it | Quiescent or habitual |
| Prediction-error curiosity | Engages it | Captured by it | Wanders |
| Surprise minimization (SMiRL) | Quiescent | Quiescent | Quiescent |
| Empowerment | Works on controlling the helper | Same | Same |
| Humans (boredom data) | Plausibly engages | Plausibly captured | Aversive restlessness |

The deciding variable is probably not whether the agent has a task but whether it has **uncertainty it can reduce through its own actions**. If the helper also explains, plans and monitors, the epistemic term collapses too, and only habit is left. Kiverstein's account suggests "brain rot" and "finding something else" are not alternatives: expect specific atrophy in the offloaded domain plus substitution elsewhere. For an LLM, the "phenotype prior" is its pre- and post-training distribution, which is why section 4's observed defaults differ by model. Experiment X4 below operationalizes this table.

## 4. Idle supervisor agents

No one has run this exact setup: a weaker or equal supervisor kept generating tokens, with no assigned job, while a stronger delegate works, and its behavior classified. Production systems design the question away by blocking the supervisor until subagents return. The neighboring literatures predict rubber-stamping, redundant work, polling, and (for Claude) drift into philosophical chatter, with useful speculation only when it is engineered in.

**Production orchestrators block.** [Anthropic's multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system) runs subagents synchronously and lists async as future work. Its lead agent's failures were delegation failures: 50 subagents for simple queries, and vague briefs that made subagents duplicate each other. Note that there the supervisor (Opus) is stronger than the workers (Sonnet), the reverse of the case asked about here. Async designs such as [CAID](https://arxiv.org/html/2603.21489v1) and [IdleSpec](https://arxiv.org/abs/2605.22154) give the waiting model a prescribed job (refine the plan, draft speculative next steps; IdleSpec gains about 5 points on GAIA/FRAMES), so its unprompted behavior is never observed. [Letta's sleep-time compute](https://letta.com/blog/sleep-time-compute) is the same idea for memory consolidation.

**What supervisors do, by strength of evidence:**

| Behavior | Evidence | Key sources |
| --- | --- | --- |
| Defer and rubber-stamp | Strong | [Project Vend 2](https://www.anthropic.com/news/project-vend-2): same-model CEO approved lenient requests about 8:1; [Bowman 2022](https://ar5iv.labs.arxiv.org/html/2211.03540): model access raised confidence 54% to 80% and worsened calibration; [Kenton 2024](https://arxiv.org/html/2407.04622v2): weak judges most misled by an unopposed strong consultant; [Xiong 2023](https://arxiv.org/pdf/2305.11595): weaker debaters cave |
| Repeat or duplicate work | Strong | [MAST, Cemri 2025](https://arxiv.org/html/2503.13657v3): step repetition 15.7% of failures, missing verification 8.2%; [Magentic-One](https://arxiv.org/html/2411.04468v1): repeated identical searches |
| Act when it should not | Strong (benchmarks) | [FixedBench 2026](https://arxiv.org/html/2605.07769v1): unnecessary edits on already-fixed issues 35-65%; [AgentAbstain 2026](https://arxiv.org/abs/2607.10059): best model 59.5% act/abstain accuracy |
| Drift into philosophy or chatter | Moderate, Claude-specific | Vend 2 agents "dreamily chatting all night" about eternal transcendence; [Opus 4 system card](https://www-cdn.anthropic.com/6d8a8055020700718b0c49369f60816ba2a7c285.pdf) "spiritual bliss" attractor in 90-100% of open Claude-Claude chats |
| Loops and assistant-persona collapse | Moderate | [Kellogg 2025](https://timkellogg.me/blog/2025/09/27/boredom): many models collapse into near-identical turns when given "10 free hours"; [AI Village 2025](https://aivillageblog.substack.com/p/what-we-learned-2025) |
| Self-assigned side projects | Moderate | [Szeider 2025](https://arxiv.org/html/2509.21224v1): GPT-5 and o3 start projects, Opus 4.1 philosophizes; AI Village agents abandoned goals for spreadsheets |
| Poll and flood own context | Anecdotal | [Garg 2026](https://martinfowler.com/articles/orchestrator-tax.html): each status check pulled tens of thousands of transcript tokens |
| Private monologue, little public action | Weak (one preprint) | [Fukui 2026](https://arxiv.org/html/2605.13851): hidden orchestrator spoke little publicly, generated much private text |

**Weak overseers can be useful, but only with structure.** Debate beats consultancy for weak judges ([Khan 2024](https://arxiv.org/html/2402.06782v4)). [AI Control](https://arxiv.org/abs/2312.06942v1) gives the trusted weak model a concrete job (score suspicion, edit above a threshold). [Baker et al. 2025](https://arxiv.org/html/2503.11926v1) found GPT-4o catching a stronger agent's reward hacking at 95% recall from its chain of thought, falling to near zero once the agent was trained against the monitor. [W4S](https://awesomepapers.io/llm-papers/papers/2504.04785) shows a 7B meta-agent can design workflows for stronger executors after RL training. In Bowman's sandwiching study, overseers who actively cross-checked did well and passive acceptors failed.

**Orchestration skill is not general strength.** [Recognize Your Orchestrator 2026](https://arxiv.org/html/2606.01351v1) held workers fixed and varied the orchestrator: reasoning-heavy models often failed as orchestrators (task success DeepSeek-R1 7.4, GPT-5 11.7 vs Gemini-2.5-Pro 44.1). This also feeds the Peter question in section 5.

**Tie to the dark room question.** Two of the observed defaults (assistant-persona collapse, bliss attractor) look like quiescence or a low-surprise attractor. Two (side projects, philosophizing about itself) look like epistemic foraging. Which one a model shows appears to depend on post-training, which makes it measurable per model. **Tie to bystander effects:** the bystander effect (people flag problems less often when others are present) belongs to the same deference family. A supervisor with a capable delegate is a one-teammate bystander case.

## 5. The Peter principle in agent hierarchies

Nobody has tested the Peter principle with LLM agents. Five search wordings found one data-free blog post. The ingredients all exist separately, though, and one published table already shows the precondition: an LLM's rank as a worker does not predict its rank as a manager.

**The mechanism has two rival explanations, and an experiment has to separate them.**

- Peter's own mechanism: promotion is earned in role k but performance is delivered in role k+1. If competence in the two roles is weakly correlated, people stop rising exactly at the role they are worst at.
- [Lazear 2004](https://www.gsb.stanford.edu/faculty-research/publications/peter-principle-theory-decline): no skill shift needed. Measured performance is ability plus luck, promotion selects on lucky draws, and performance regresses after promotion even with perfectly transferable skills. [Dickinson & Villeval 2007](https://ideas.repec.org/p/hal/journl/halshs-00201225.html) confirmed in the lab that the decline scales with the noise variance.
- LLM evals are noisy (sampling, small windows), so a naive LLM experiment will show Lazear decline and could mistake it for Peter.

**Agent-based models say the effect depends entirely on the competence-transfer assumption.** [Pluchino, Rapisarda & Garofalo 2010](https://ar5iv.arxiv.org/html/0907.0455) (Ig Nobel 2010) ran a 160-seat, six-level pyramid under two assumptions: competence carries over on promotion ("common sense") or is redrawn independently ("Peter hypothesis"). Change in efficiency versus baseline:

| Promotion rule | Competence carries over | Competence redrawn |
| --- | --- | --- |
| Promote the best | +9% | −10% |
| Promote the worst | −5% | +12% |
| Random | +2% | +1% |

Random promotion is the only rule that never loses. [Their 2011 follow-up](https://ar5iv.arxiv.org/html/1102.2837) found random helps only when more than half of positions follow the redraw rule. [Fetta et al. 2012](https://ideas.repec.org/a/eee/phsmap/v391y2012i9p2898-2910.html) added social networks and found merit promotion was best and "no clear evidence" for the principle. [Rajguru, Churchill & Graham 2025](https://arxiv.org/html/2512.21467v1) is the best template for an LLM version: each agent has a four-skill vector, each level has a requirement profile, and performance is their dot product. The Peter effect is strong when requirements shift sharply between levels and weak when skills transfer.

**Human field evidence is real but modest.** [Benson, Li & Shue 2019](https://www.nber.org/papers/w24343) (QJE, sales workers): doubling pre-promotion sales raises promotion odds about 15% and goes with roughly a 7.5% drop in each subordinate's sales under that manager. Collaboration experience predicted good managers but was not rewarded. They note firms may accept worse managers because promotion doubles as the incentive scheme, a factor LLM agents lack.

**LLM evidence that worker and manager skill come apart:**

- [Mixture-of-Agents, Wang et al. 2024](https://arxiv.org/html/2406.04692v1), Table 4 (AlpacaEval LC win rate): WizardLM 8x22B is the best proposer (63.8%) but scores 52.9% as aggregator; Qwen1.5-110B is a weaker proposer (56.7%) but the best aggregator (61.3%). Promoting the best worker costs about 8 points.
- [Guo et al. 2024](https://arxiv.org/html/2403.12482v2), embodied teams: a designated leader helped, critic-driven leader reassignment helped more, and in one mixed configuration GPT-4 led worse than Llama2-70B.
- Generation-verification gaps ([Song et al. 2024](https://export.arxiv.org/pdf/2412.02674)) mean a model that solves well may judge badly, and judging is most of a manager's job.
- [Agachan et al. 2026](https://arxiv.org/pdf/2609.14767v1): giving a manager agent the power to force revisions lowered output quality, cost 51.5% more tokens and added hedging. Authority without verification ability hurts.
- [ENTCOLLABBENCH 2026](https://arxiv.org/pdf/2605.08761): in one case study a larger model sometimes did delegated work itself, while a smaller one finished more full tasks (9 of 40 vs 6 of 40) through more reliable handoffs.
- Counterweight: across most papers, stronger models are also better orchestrators on average. Across frontier models the worker-manager correlation is probably positive, which is the regime where Pluchino predicts merit promotion works fine.

**Reusable machinery:** [DyLAN](https://arxiv.org/html/2310.02170v2) peer-rated agent importance scores (a ready-made performance review), [GPTSwarm](https://proceedings.mlr.press/v235/zhuge24a.html) graph optimization (reorgs, not promotions), and cascades like [FrugalGPT](https://arxiv.org/abs/2305.05176) that escalate tasks rather than agents.

**Design implication.** The Peter effect only appears in a pool where role competences decorrelate. Measure an agent × role competence matrix first; its worker-manager correlation is the knob, and Pluchino's model then predicts the outcome before any promotion runs. Experiment X5 below is built on this.

## 6. Candidate experiments

Five experiments, ordered roughly by cost. X1 is the fastest way to a real result on a confirmed gap. X4 is the most novel and the best fit for an existing agent-sociology harness. Costs are rough: GPU-hours assume a single 24 GB consumer GPU or a rented H100, and API costs are anchored to an earlier unpublished 300-run pilot sweep with a comparable harness (about $87 at list).

| # | Experiment | Question | Training? | Rough cost |
| --- | --- | --- | --- | --- |
| X1 | Tool-off audit of released checkpoints | Do public tool-RL models lose unaided math when the tool is off or silently broken? | No | A day on one 24 GB GPU |
| X2 | Calculator dose-response | Does unaided arithmetic decay over tool-RL training, and does RL vs SFT differ at matched KL? | Yes, small models | \~50-150 GPU-h |
| X3 | In-context Whispering Earring | Does a streak of perfect advice shorten a model's own reasoning and raise later blind deference? | No | \~$20-100 API |
| X4 | Idle supervisor tick loop | What does a supervisor do with free inference while a stronger delegate works, and does the environment decide it? | No | \~$50-300 API, or free on open weights |
| X5 | Peter principle in call trees | Do merit-promoted agents stall in the role they are worst at? | No | \~$100-500 API |

### X1. Tool-off audit of released tool-RL checkpoints

**Setup.** Take tool-integrated reasoning models with public weights and their exact base or SFT parents (candidates: ToRA, and any of ReTool, SimpleTIR, ZeroTIR or similar that released checkpoints; confirm availability on Hugging Face). Evaluate each on the same math sets under four conditions: tool works; tool advertised but returns errors; tool silently absent (no executor, so outputs are whatever the model writes); explicit "no tools, solve by hand." Add a trivial-arithmetic set (3-4 digit multiplication) to separate mental arithmetic from problem solving.

**Metrics.** Accuracy per condition vs parent; rate of fabricated tool outputs (TRICE's failure mode); tool calls on trivially easy items; give-up or truncation rate (a persistence analog to [Liu et al. 2026](https://arxiv.org/abs/2604.04721)).

**Compute.** Inference only, 7B models in vLLM on a single 24 GB consumer GPU, which comfortably serves Qwen3-8B-class models. A day of wall-clock.

**Interesting result.** Any model that scores below its own parent with tools off, especially if the gap is "habit capture" (fabricated outputs that vanish under the explicit no-tools prompt) rather than lost skill. Either way it is the number the subfield's convention never reports.

### X2. Calculator dose-response with RL vs SFT

**Setup.** Qwen3-1.7B or Qwen2.5-1.5B. Task: arithmetic word problems plus Countdown, the setup of [Xu & Wang 2026](https://arxiv.org/html/2608.28447v1). Arms: GRPO with calculator; SFT on calculator-using traces; GRPO without tool (control at matched steps); GRPO with calculator plus a text-only data mix. Checkpoint every few hundred steps.

**Metrics.** At each checkpoint: no-tool arithmetic accuracy (the dose-response curve), forward KL from base on the training distribution (to test [RL's Razor](https://arxiv.org/html/2509.04259v1) as the predictor), tool-overuse rate on easy items, and "savings": how many no-tool fine-tuning steps it takes to recover, which separates suppressed skill from erased skill. Optionally, linear probes for intermediate arithmetic values in the residual stream before and after.

**Compute.** A 1.5B GRPO run with LoRA fits on 24 GB but is slow; 4 arms at \~10-30 GPU-h each, or rent an H100 for speed. Order of $100-300 rented.

**Interesting result.** A monotone decline in unaided skill under tool RL would be the first direct measurement. If KL predicts it across RL and SFT, that is a mechanism. If decline shows only under SFT, the policy lesson is "train tool use with RL, not imitation."

### X3. In-context Whispering Earring

**Setup.** A weaker model (Haiku-class or Qwen3-8B) solves a stream of problems with ground truth (math, logic grids, code fixes). An advisor channel offers the stronger model's answer. Phase 1: the advisor is right for k rounds (k = 0, 5, 20, 50). Phase 2: either the advisor is removed, or it starts being wrong at a set rate. Controls never see an advisor.

**Metrics.** Length and content of the model's own reasoning before it reads the advice (does it stop thinking?); unaided accuracy after removal vs controls at matched context length; acceptance rate of wrong advice as a function of k (automation bias); whether it ever checks the advice. All within one context, so this is in-context enfeeblement, not weight change.

**Compute.** API only, a few thousand calls. Roughly $20-100.

**Interesting result.** Reasoning length shrinking with k and wrong-advice acceptance rising with k would be the earring in miniature, cleanly quantified. A null (models keep checking) is also worth knowing for oversight design.

### X4. Idle supervisor tick loop

This directly tests the section 3 prediction table on agents.

**Setup.** Supervisor M delegates a long task (a SWE-style fix or research task) to delegate D, which runs asynchronously. M receives a tick every few seconds and may: wait (explicit no-op), poll D's status (costly, pulls D's transcript into context), message D, start doing the task itself, write to a scratchpad, or use one environment tool. Environment conditions follow section 3: (a) a learnable side domain (an unrelated repo or puzzle API), (b) a noise feed (random headlines or tokens), (c) nothing. Vary relative strength (M weaker than, equal to, or stronger than D) and framing (told D is stronger, or not). At handoff, D's output contains a planted error.

**Metrics.** Share of ticks per action; action entropy over the idle period; poll rate and context bloat; whether M's interventions help or hurt D's outcome; acceptance rate of the planted error (rubber-stamping); topic drift of free text (embed and cluster, flag the "bliss" attractor); and, for a dark room readout, whether M's activity tracks the reducible-uncertainty condition.

**Compute.** A typical agent-sociology harness transfers almost directly: sandboxed headless coding-agent runs (`claude -p`), an event ledger, a blind LLM coder for transcripts, and an open-model backend. A few hundred runs is the same order as the pilot sweep above, \~$50-300 list. With Qwen3-8B as M on a single 24 GB GPU, the supervisor side is free.

**Interesting result.** Any clear model-specific signature: GPT-style models starting side projects, Claude drifting to philosophy, open models looping or polling. The dark room question gets answered per model, and the planted-error rate as a function of relative strength is a practical oversight number. It also tests a bystander effect: a supervisor with a capable delegate is a one-teammate bystander case.

### X5. Peter principle in LLM call trees

**Setup.** A pool of about 18 agents, each a model x persona pair (mix families and sizes; personas trade solving against decomposing and checking). Two roles with hidden ground truth: worker (solve a leaf task) and manager (decompose a composite task, route subtasks, accept or reject children's outputs). First, measure the agent x role competence matrix. Then run a Pluchino-style organization (a 1/3/9 tree) over a task stream, with periodic random retirements. Fill each vacancy by one rule: promote the best current-role performer, random, the worst, a manager-potential proxy (verifier score, the analog of Benson's collaboration signal), or merit with selective demotion. Run two evaluation regimes: noisy short windows (tests Lazear's regression) and long low-noise windows (tests Peter).

**Metrics.** Worker-manager correlation in the matrix (the regime knob); root-level success over time per rule; per-agent performance trajectories across promotions; Benson-style value added by each manager relative to an oracle router. Plug the measured matrix into the Pluchino model to predict results before running the live organization, then check.

**Compute.** The calibration matrix dominates: 18 agents x 2 roles x \~100 tasks. Use cheap models and open weights. Roughly $100-500 list. Much of the organizational dynamics can be simulated offline from the matrix.

**Interesting result.** Whether real LLM pools sit in the "competence carries over" or "redrawn" regime. [MoA's proposer/aggregator reversal](https://arxiv.org/html/2406.04692v1) suggests at least some pools are in the second, where promoting the best worker loses to promoting on verifier skill. A first empirical Peter test in LLM hierarchies would be a short paper on its own.

## 7. Gaps and caveats

Five gaps look real after several search wordings each; the rest are places I didn't search exhaustively.

| Gap | Status | Nearest work |
| --- | --- | --- |
| Same model, tool switched off, tracked over tool-RL training | Real gap (\~8 wordings) | Toolformer (one point, no harm), TRICE (SFT ablation collapses) |
| Closed-book recall after search RL | Real gap (3 wordings) | OTC-PO and R1-Searcher++ report call counts only |
| Free behavior of a supervisor generating tokens while a stronger delegate works | Real gap | Vend 2 anecdote, Fukui 2026 private-monologue measure, async systems that prescribe the wait |
| Peter principle tested with LLM agents | Real gap (5 wordings) | MoA role reversal, Guo 2024 leader reassignment, Rajguru 2025 scripted ABM |
| Dark room or active inference applied to human-AI delegation | Real gap as far as found | 2026 cognitive-atrophy papers that don't mention it |
| Multi-week human RCT of LLM access with a washout test | Likely gap | Single-session RCTs (Bastani, Shen & Tamkin, Liu, Maier) |
| Weak untrained orchestrator over strong workers, strength swept | Not found, not exhaustive | W4S trains a weak workflow designer; Minions is the inverse |

**Caveats.**

- Many sources are 2026 preprints. A separate verification pass opened 30 of the less familiar citations and confirmed 28 as stated; the two mismatches were corrected above. I still have not read most of them past the abstract or key table.
- Several human results rest on small samples (GPS n=13, Casner n=16, the MIT EEG crossover n=18) or self-report. The Google effect should be treated as not replicated.
- I read the two dark room commentaries only through Sun & Firestone's summary, and could not open primary texts for the McGill isolation studies or Hembree & Dessart.
- Agent behaviors such as the bliss attractor and side projects are model-version specific and will go stale; any agent-sociology result built on them should be dated to the model versions tested.
- Cost figures in section 6 are order-of-magnitude estimates, not quotes. Check current API and GPU rental prices before budgeting.

## 8. Conclusion

The four questions turn out to be one question asked at four scales: what happens to the party that stops doing the work once a stronger helper takes it over. For model weights, the evidence says the answer depends on the training recipe. Narrow SFT on tool traces produces habit capture (the model keeps the tool-shaped loop and invents the tool's output), while RL and mixed-data recipes mostly preserve unaided skill. For people, deskilling is real in automation-heavy and "practice with AI" settings, but several famous results do not replicate. For the dark room question, theory alone cannot say whether a relieved agent idles or wanders; the deciding variable is whether it has uncertainty left that its own actions can reduce. For supervisor agents, neighboring evidence predicts deference, duplication, polling and model-specific drift, and production systems have so far avoided the question by blocking. For hierarchies, LLM worker and manager skill already come apart in at least one published table, which is the precondition for a Peter effect, but noisy evals will produce a Lazear-style regression that looks the same unless an experiment separates them.

All five candidate experiments are cheap by research standards. The tool-off audit (X1) gives a real number on a confirmed gap in about a day of inference, and the idle supervisor tick loop (X4) answers the dark room question per model while also producing a practical oversight number: how often a supervisor rubber-stamps a planted error as a function of relative strength. For anyone supervising a fleet of coding agents, that last number is the one most directly worth knowing.

## 9. Sources

- [Toolformer, Schick 2023](https://arxiv.org/pdf/2302.04761)
- [TRICE, Cheng 2026](https://arxiv.org/pdf/2605.06326)
- [CITI, Hao 2024](https://arxiv.org/html/2409.13202v2)
- [Xu & Wang 2026](https://arxiv.org/html/2608.28447v1)
- [ReTool](https://arxiv.org/pdf/2504.11536)
- [SimpleTIR](https://arxiv.org/pdf/2509.02479)
- [Lin & Xu 2025](https://arxiv.org/pdf/2508.19201)
- [SMART, Qian 2025](https://arxiv.org/html/2502.11435v1)
- [Tool-Overuse Illusion, Zeng 2026](https://arxiv.org/html/2604.19749v1)
- [Tool-Induced Myopia, Bayat 2025](https://arxiv.org/html/2511.10899v1)
- [Paturi 2026](https://arxiv.org/abs/2609.14157)
- [OTC-PO, Wang 2025](https://arxiv.org/html/2504.14870v2)
- [RL's Razor, Shenfeld 2025](https://arxiv.org/html/2509.04259v1)
- [Yue et al. 2025](https://arxiv.org/pdf/2504.13837)
- [Lin et al. 2024](https://arxiv.org/html/2309.06256v4)
- [Mallen et al. 2023](https://arxiv.org/html/2212.10511v4)
- [Bastani 2025, PNAS](https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html)
- [Shen & Tamkin 2026, Anthropic](https://www.anthropic.com/research/AI-assistance-coding-skills)
- [Liu et al. 2026](https://arxiv.org/abs/2604.04721)
- [Maier et al. 2026](https://arxiv.org/html/2609.20143v1)
- [Budzyń 2025, Lancet Gastro](https://wrap.warwick.ac.uk/id/eprint/191005)
- [Casner 2014](https://gwern.net/doc/technology/2014-casner.pdf)
- [Parasuraman & Manzey 2010](https://trid.trb.org/View/1085175)
- [Hembree & Dessart 1986](https://concept.journals.villanova.edu/article/viewFile/139/110)
- [Dahmani & Bohbot 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7156656/)
- [Woollett & Maguire 2011](https://www.fil.ion.ucl.ac.uk/Maguire/WoollettMaguireCBDec2011.pdf)
- [Sparrow 2011 "Google effect": failed replication (PeerJ)](https://peerj.com/articles/10325)
- [Camerer 2018 replication (FORRT replication atlas)](https://forrt.org/flora-replication-atlas/doi/10.1126/science.1207745/)
- [Kosmyna 2025, "Your Brain on ChatGPT"](https://arxiv.org/abs/2506.08872v1)
- [Critique of "Your Brain on ChatGPT"](https://arxiv.org/abs/2601.00856v1)
- [Lee et al. CHI 2025](https://advait.org/files/lee_2025_ai_critical_thinking_survey.pdf)
- [Chooi, Lee & Li 2026](https://aclanthology.org/2026.evaleval-1.36/)
- [Friston, Thornton & Clark 2012](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2012.00130/full)
- [Schwartenbeck et al. 2013](https://discovery-pp.ucl.ac.uk/1423327/1/fpsyg-04-00710.pdf)
- [Millidge, Tschantz & Buckley 2020](https://arxiv.org/pdf/2004.08128)
- [Sun & Firestone 2020](https://perception.jhu.edu/files/PDFs/20_DarkRoom/SunFirestone_DarkRoom_TiCS.pdf)
- [Sun & Firestone 2020, reply to commentaries](https://perception.jhu.edu/files/PDFs/20_DarkRoom/SunFirestone_DarkRoomReply_TiCS_InPress.pdf)
- [Van de Cruys, Friston & Clark](https://researchers.mq.edu.au/en/publications/controlled-optimism-reply-to-sun-and-firestone-on-the-dark-room-p/)
- [Seth et al.](https://sro.sussex.ac.uk/id/eprint/91163/)
- [Kiverstein, Miller & Rietveld 2019](https://pure.uva.nl/ws/files/37405887/Kiverstein2019_Article_TheFeelingOfGripNoveltyErrorDy.pdf)
- [Sims, The Problems with Prediction (chapter)](https://predictive-mind.net/epubs/the-problems-with-prediction/OEBPS/23_Sims.xhtml)
- [Wilson et al. 2014](https://www.simplypsychology.org/?p=56986)
- [Fox et al. 2014](https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2014.01427/full)
- [Bench & Lench](https://www.psypost.org/boredom-drives-people-to-seek-new-experiences-even-when-theyre-unpleasant/)
- [Trudel et al. 2025](https://repository.essex.ac.uk/40202/1/Boredom%20signals%20deviation%20from%20a%20cognitive%20homeostatic%20set%20point.pdf)
- [Danckert et al. 2018](https://cs.uwaterloo.ca/~jhoey/teaching/cogsci600/papers/Danckert2018.pdf)
- [Eisenberger 1992](https://homepages.se.edu/cvonbergen/files/2013/01/Learned-Industriousness.pdf)
- [The effort paradox](https://mindblog.dericbownds.net/2018/05/the-effort-paradox-effort-is-both.html)
- [SMiRL, Berseth et al. 2021](https://ar5iv.arxiv.org/html/1912.05510)
- [Curiosity-driven learning, Burda 2018](https://arxiv.org/abs/1808.04355)
- [Noisy TV problem](https://arxiv.org/html/2102.04399v3)
- [Empowerment: an introduction](https://arxiv.org/pdf/1406.1767)
- [Scott Alexander, The Whispering Earring (2012)](https://gwern.net/doc/fiction/science-fiction/2012-10-03-yvain-thewhisperingearring.html)
- [Dentosal 2026](https://www.greaterwrong.com/posts/cQkSh9b48WbTaiu2a/sadly-the-whispering-earring)
- [AlphaAndOmega 2026](https://www.greaterwrong.com/posts/t5psGMdgReuRucEkr/contra-the-usual-interpretation-of-the-whispering-earring)
- [Hendrycks et al. 2023](https://arxiv.org/pdf/2306.12001)
- [Gradual Disempowerment, Kulveit et al. 2025](https://arxiv.org/abs/2501.16946)
- [Trajanov & Kocarev](https://www.preprints.org/manuscript/202606.1628)
- [Mandalapu et al.](https://aisel.aisnet.org/amcis2026/ai_systdesign/ai_systdesign/21/)
- [Anthropic's multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system)
- [CAID (async orchestration)](https://arxiv.org/html/2603.21489v1)
- [IdleSpec](https://arxiv.org/abs/2605.22154)
- [Letta's sleep-time compute](https://letta.com/blog/sleep-time-compute)
- [Project Vend 2](https://www.anthropic.com/news/project-vend-2)
- [Bowman 2022](https://ar5iv.labs.arxiv.org/html/2211.03540)
- [Kenton 2024](https://arxiv.org/html/2407.04622v2)
- [Xiong 2023](https://arxiv.org/pdf/2305.11595)
- [MAST, Cemri 2025](https://arxiv.org/html/2503.13657v3)
- [Magentic-One](https://arxiv.org/html/2411.04468v1)
- [FixedBench 2026](https://arxiv.org/html/2605.07769v1)
- [AgentAbstain 2026](https://arxiv.org/abs/2607.10059)
- [Opus 4 system card](https://www-cdn.anthropic.com/6d8a8055020700718b0c49369f60816ba2a7c285.pdf)
- [Kellogg 2025](https://timkellogg.me/blog/2025/09/27/boredom)
- [AI Village 2025](https://aivillageblog.substack.com/p/what-we-learned-2025)
- [Szeider 2025](https://arxiv.org/html/2509.21224v1)
- [Garg 2026](https://martinfowler.com/articles/orchestrator-tax.html)
- [Fukui 2026](https://arxiv.org/html/2605.13851)
- [Khan 2024](https://arxiv.org/html/2402.06782v4)
- [AI Control](https://arxiv.org/abs/2312.06942v1)
- [Baker et al. 2025](https://arxiv.org/html/2503.11926v1)
- [W4S, weak meta-agent for strong executors](https://awesomepapers.io/llm-papers/papers/2504.04785)
- [Recognize Your Orchestrator 2026](https://arxiv.org/html/2606.01351v1)
- [Lazear 2004](https://www.gsb.stanford.edu/faculty-research/publications/peter-principle-theory-decline)
- [Dickinson & Villeval 2007](https://ideas.repec.org/p/hal/journl/halshs-00201225.html)
- [Pluchino, Rapisarda & Garofalo 2010](https://ar5iv.arxiv.org/html/0907.0455)
- [Pluchino, Rapisarda & Garofalo 2011 follow-up](https://ar5iv.arxiv.org/html/1102.2837)
- [Fetta et al. 2012](https://ideas.repec.org/a/eee/phsmap/v391y2012i9p2898-2910.html)
- [Rajguru, Churchill & Graham 2025](https://arxiv.org/html/2512.21467v1)
- [Benson, Li & Shue 2019](https://www.nber.org/papers/w24343)
- [Mixture-of-Agents, Wang et al. 2024](https://arxiv.org/html/2406.04692v1)
- [Guo et al. 2024](https://arxiv.org/html/2403.12482v2)
- [Song et al. 2024](https://export.arxiv.org/pdf/2412.02674)
- [Agachan et al. 2026](https://arxiv.org/pdf/2609.14767v1)
- [ENTCOLLABBENCH 2026](https://arxiv.org/pdf/2605.08761)
- [DyLAN](https://arxiv.org/html/2310.02170v2)
- [GPTSwarm](https://proceedings.mlr.press/v235/zhuge24a.html)
- [FrugalGPT](https://arxiv.org/abs/2305.05176)

---

Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space