Agent0: A self-evolving multi-agent curriculum without human seed data¶

Deterministic core mechanism, not the original benchmark.

Paper information¶

Field	Details
Paper	Agent0 (arXiv 2511.16043)
Institution	University of North Carolina at Chapel Hill / Salesforce Research / Stanford University.
First public release	2025-11-20 (arXiv v1)
Author code	No official author code found or released.
Local adapter or method key	`agent0`
Local reproduction code	`src/auto_research/agent_research/`

Paper method and evidence¶

A proposer generates verifiable tool tasks. Multiple execution agents produce candidates and vote; increase curriculum difficulty along the current capability boundary.

flowchart LR
  T[Propose verifiable task] --> A[Multiple execution agents]
  A --> V[Majority vote and verifier]
  V --> C[Advance curriculum]

Original paper figure¶

The original figure shows task proposal and multi-agent verification. From the original paper; copyright remains with the authors.

Core equation¶

\[ x\sim\pi_{task},\ y_{1:K}\sim\pi_{agent},\ \hat y=\operatorname{mode}(y_{1:K}),\quad R=V(x,\hat y). \]

Reported results¶

The paper bootstraps tool-integrated reasoning without seed data and reports improvements across agent benchmarks. No production A/B.

Local reproduction¶

PlanBench mini, 120 episodes, seed 42: joint success 1.0000, mean cost 0.3200, and nonzero method telemetry.

auto-research agent-eval --method agent0 --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "Combine agent0 with installed paper operators" --generations 2 --population 4

Fixed-seed metrics. Only the method-specific objective, state updates and matched budget are tested. No large model, multi-GPU RL, private environment, live web or full benchmark.