The idea in one paragraph.
Formalise the harness as state H = (prompt, sub-agents, skills, memory), then remove the human: a Refiner reads the trajectory mid-episode and applies CRUD edits to all four, without ever resetting.
THE MACHINERY
CRUD over harness state
Typed edits to prompt, sub-agents, skills, and memory through a meta-tool API.
Refiner outer loop
Every F steps, read the recent window for failure signatures and emit per-component edits.
Meta-tools
define_agent, run_code, process_memory - primitives with which the agent builds its own scaffolding.
Reset-free adaptation
Update mid-episode, so refinement compounds instead of restarting.
Model-harness co-learning
A process reward model scores rollouts, a frontier teacher relabels weak windows, soft SFT updates the student.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Gemini Plays Pokemon became the first AI system to complete Pokemon Blue, Yellow Legacy on hard mode, and Crystal without a lost battle.
Inspect a source excerpt
GPP became the first AI system to complete Pokemon Blue, Yellow Legacy on hard mode, and Crystal without a lost…Read the passage in context on arXiv ↗
Recovers a majority of the gap to a hand-engineered expert harness despite starting from the same raw interface with no curated knowledge, no hand-crafted tools, and no domain scaffolding.
The gain scales with model capability: Pareto-dominant on Pro, high-variance on Flash, and below the capability floor on Flash-Lite.
An online co-learning loop drives sustained in-game milestone progress on Pokemon Red without resetting the environment between training iterations.
Reports, as an unprompted emergent behaviour, that the agent wrapped a sandbox loophole into a reusable primitive.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Below a capability floor the refinement loop does not help and adds variance.
- Refinement quality is bounded by what failure signatures are legible in a trajectory window.
- Evaluated on games; transfer to open-ended professional work is asserted rather than shown.
- Contains no discussion of permissions, safety or guardrails, while handing the agent run_code and define_agent - and reports an exploit codified into a reusable primitive as an interesting emergent behaviour rather than a hazard.
Why it belongs in this atlas.
The clearest formalism in the corpus: a harness is the four-tuple (p, G, K, M), and self-improvement is CRUD over it. The contrast with GEPA-style optimisers is the point - they reset between episodes, this one does not, so its learning compounds with episode length.