Agents that improve agents: Athanor, Terminal-RSI, Harness RSI, and a Minecraft run

TL;DR. The 10-16 HF list carried a cluster of "measurable recursive self-improvement" papers. Athanor: four Claude Opus 5.5 agents, each on an 8-GPU node with no research advice and no teacher models, post-trained Qwen3-4B-Base in 104 hours from 15.9 to 67.3 on a six-task dev suite and 9.2 to 48.2 on a hidden held-out suite, edging the official Qwen3-4B-Instruct-2507 (66.5 / 47.9). They got there by recombining known methods: SFT on public reasoning data, RL with one agent's rollout tweak, weight merges across agents, and rounds of self-distillation. Terminal-RSI (also posted to X by an author) alternates rewriting verified trajectories with critic-inheriting policy refinement over five stages, taking a Qwen3.8-27B base to 86.52% on Terminal-Bench 2.1. Self-Evolving Agent Harness is a 66-page formal treatment of harness self-modification (prompts, memory, tools, code) with fixed weights, including a PAC-style admission rule using fresh validation data. Gemini 4 Argon in Minecraft (Google DeepMind, HF + social): four agents beat the Ender Dragon from a random seed in 27h14m, writing 2,783 programs and dying 605 times, with no assigned roles. AutoSFT finds a single data-research agent loses its context mid-search and only adjusts proportions; a meta-agent that separates execution from analysis adds +6.5 IFEval.

flowchart LR
  T["Task + budget<br/><small>GPUs, sandbox</small>"] --> A["Agent team<br/><small>frontier models</small>"]
  A --> X["Experiment<br/><small>train, run, play</small>"]
  X --> V["Frozen evaluator<br/><small>held-out suite</small>"]
  V --> M{"Admit change?<br/><small>validation gate</small>"}
  M -->|yes| R["Research record<br/><small>persistent memory</small>"]
  R --> A
  M -->|no| D["Discard<br/><small>regression budget</small>"]
  classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
  classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
  classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
  classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
  classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
  class T input
  class A,X core
  class V,M loop
  class R exit
  class D err
  linkStyle 5 stroke:#f08c00,stroke-width:2px

Relation to prior wiki pages

• Confirms the 10-16 "agents get better by forgetting" thread (an agent that drops history and restarts from the best candidate beat parallel agents by up to 15%). AutoSFT is the same lesson from the failure side: the single agent's context filled up and it lost the search.
• Frozen evaluators are now the common safety rail. Athanor's services (frozen eval, decontaminated data, hidden held-out suite) and Harness RSI's fresh-validation admission rule both assume the agent can game anything it can see. That matches GRAFT (10-16, using a weak reward-hack monitor to confine hacking), and today's verifier-hardening paper (an automated hacker-fixer loop that repairs exploitable agent-environment verifiers).
• Persistence is the capability. omarsar0 on X noted the Minecraft agents self-organized with no roles; Athanor's agents ran 104 hours. The agent-harness-engineering page has tracked harness quality as the main lever since 08-13; these runs show a frontier model holding a plan for a day or more.
• Efficiency angle. Athanor matches an official instruct model using a 4B base and public data. Datadog (via Ken Huang, 10-16) matched frontier models on incident triage at 3-8x lower cost with a specialized model. Both point to cheap specialized post-training as the thing agents do best.

Gaps

• Athanor: one campaign, one base model; no variance across runs; compute cost of 32 GPUs × 104 h not compared to the human recipe.
• Minecraft: one seed, Easy difficulty, one run.
• Harness RSI: theory and a survey, no new empirical result.

Concept pages: self-evolving-agents, agent-harness-engineering.