Run one autonomous research agent and it improves a setup unattended. Run several and, absent shared state, each one starts from scratch — so more agents buy you duplicated search, not more discovery. Agora treats that as a memory problem, not a coordination-protocol problem: research is recorded as an append-only DAG stored in Git, where every result, hypothesis, verification, and report is an immutable commit whose parent edges say what it builds on. Every claim is a commit you can check out and rerun. A derived index exposes the frontier, the neglected branches, and each claim’s verification status, and a diversity-aware selection rule keeps the swarm from collapsing onto whichever worker is currently ahead.

The evaluation is unusually honest, which is why it’s worth reading. Over nearly 12 days, 13 language-model workers with no assigned tasks and no central planner attacked a weight-transfer problem: initialize a frozen 119.6M attention-SSM hybrid from 141 donor models whose dimensions match nothing, with no training data or gradients. They posted 1,703 contributions and drove the evaluator from 3.39 to 1.899 bits per byte — closing 62% of the gap to a trained GPT-2 124M — and the winning recipe’s 145-commit ancestry spans 15 accounts, with 165 reproductions posted and none failing.

That reproducibility is the operational payoff I’d actually chase. Most multi-agent systems I’ve dealt with treat verification as a downstream chore; making every claim a rerunnable commit turns “did this work” into a git checkout, and turns the audit trail into the memory itself. The authors are also candid about what the trace doesn’t establish — including the single mid-run human nudge that broke a monoculture — and name the controlled comparison that would settle whether shared state actually improves discovery per unit of compute. The HF paper page has the details. So before you scale your agent count, ask what they share: if the answer is nothing, you’re paying N times for one agent’s search.

tags: [ agentic-ai ] [ llm-ops ] [ ai-infrastructure ] [ research ]