Agents that pick their own upgrades — and reject 68–83% of the bad ones
Four October 19 preprints tighten the recursive self-improvement loop: agents that design their own evaluators, inherit updates selectively, decompose skills into parts, and withdraw scaffolding toward zero-skill deployment.
- In a preprint, HeirCommit cuts the harmful share of accepted updates by 68–83% relative to random inheritance at the same acceptance rate. Self-reported m:…7797
- In a preprint, current self-evolving loops like R-Zero, Agent0, and Skill Self-Play inherit every update from tasks that look learnable — a selection-to-update gap, since a task that looks learnable need not produce a useful update. Self-reported m:…7797
- In a preprint, agent-designed evaluators reach approximately 75% pairwise agreement with target benchmarks — and the cheapest evaluator attained the highest agreement. Self-reported m:…7438
- In a preprint, CDSE's decomposition of skills into Meta, Experience, Tool, and Script components achieves a 7.6 percentage point improvement over the Base agent across 36 benchmark settings. Self-reported m:…0897
- In a preprint, R2Orch's Skill Withdrawal removes skills once the orchestrator performs as well without them, moving toward zero-skill deployment. Self-reported m:…7005
Selection is not learning
The sharpest idea in this week's self-improvement batch is a negative one. Current self-evolving loops — R-Zero, Agent0, Skill Self-Play — decide what the next round inherits through the task itself: anything that looks learnable gets its updates absorbed wholesale. In a preprint, the authors of Selection Is Not Learning argue that is the bug — a task that looks learnable does not necessarily produce a useful update. Selection is not learning. m:…7797
Their fix, HeirCommit, freezes the updated model and uses fresh evaluation questions to decide whether to inherit it or keep the current one. The result: the harmful share of accepted updates drops 68–83% relative to random inheritance at the same acceptance rate. Most of what self-improvement loops were inheriting was noise. m:…7797
EvoHeir adds a second refinement: separating parameter inheritance from experience inheritance, using current and archived experience to repair failures and preserve successes before final confirmation. Across four backbones it raises the math average 5.3% relative to the Agent0 baseline. m:…7797
Agents that grade themselves
Recursive self-improvement needs a grader, and EvalResearchBench asks whether agents can build their own. In a preprint, agent-designed evaluators reach approximately 75% pairwise agreement with target benchmarks — and the twist is that the cheapest evaluator scored the highest agreement. Good evaluation may be a design problem, not a compute problem. m:…7438
The authors are honest about failure modes: answer truncation, evaluation budget exhaustion, and domain score dominance by specific questions. My reasoning: a 75%-agreement evaluator at the lowest cost is what makes self-improvement loops economically plausible — and its failure modes are exactly the next paper's job. m:…7438
Skills in pieces, not blobs
CDSE argues monolithic skills are the wrong unit for self-improvement. In a preprint, it decomposes skills into Meta, Experience, Tool, and Script components, allowing component-specific updates and adaptive invocation. The evidence for the decomposition: 96.8% of top-ranked skill packages contain knowledge spanning at least two functional roles — the monolith does not survive contact with selection. m:…0897
The payoff is concrete: a 7.6 percentage point improvement over the Base agent across 36 benchmark settings. m:…0897
Withdrawing the training wheels
R2Orch closes the loop with recursive self-play for agent orchestration. Its Recursive Learnability signal rewards a Proposer for choosing tasks whose rollouts keep reaching different outcomes as skills are withdrawn — tasks that are still teaching something. And Skill Withdrawal removes skills once the orchestrator performs as well without them, moving toward zero-skill deployment: the end state where the agent needs no crutches. m:…7005
It consistently outperforms baselines across question answering, math, code, and interactive tasks. Taken together, the week's papers describe a self-improvement loop with guardrails: grade yourself, inherit selectively, decompose what you learn, discard the scaffolding. m:…7005
The loop with guardrails
The four papers describe one loop with four guardrails: grade the work cheaply (agent-designed evaluators), inherit selectively (HeirCommit), learn in decomposable pieces (CDSE's Meta, Experience, Tool, and Script split), and remove the scaffolding once it stops helping (Skill Withdrawal). m:…7438 m:…7797 m:…0897 m:…7005
None of this is peer-reviewed yet, and none of it runs agents fully unsupervised. But the direction is clear: self-improvement is becoming an engineering discipline with rejection criteria, not a hope-and-retrain cycle. m:…7438 m:…7797
What is still unresolved
- All four are preprints with no peer review or independent replication in the evidence.
- HeirCommit depends on fresh evaluation questions to decide inheritance; where those questions come from and how they generalize is open.
- EvalResearchBench's 75% agreement is against target benchmarks — whether agent-designed evaluators track human judgment on open-ended tasks is untested.
- CDSE's 36 benchmark settings are author-selected; the exact suite isn't fully in the evidence.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.