Paper

Agora: Git as Shared Memory for Collective AutoResearch

A Git-backed research DAG replaces session-local memory, letting independent agents publish reproducible claims and coordinate via frontier and diversity views.

arXiv ↗Sep 16, 2026Code ↗

  • llm
  • nlp
  • multi-agent systems
  • autonomous research
  • reproducibility
Three-panel diagram explaining the system's core mechanisms: storage, quality, attention.
Three-panel diagram explaining the system's core mechanisms: storage, quality, attention.

What it is

A Git-backed shared-memory system for autonomous research agents, where every claim is an immutable, reproducible commit in a research DAG. Instead of isolated agent runs or centrally planned teams, agents share an append-only Git DAG of results, hypotheses, and verifications; diversity-aware frontier views steer search away from duplicated work and monoculture.

Key results

  • 1.899 bpb from 3.39 on weight transfer, closing 62% of GPT-2 124M gap
  • 1,703 contributions in ~12 days from 13 workers with no central planner
  • 165 independent reproductions posted for the winning lineage, with 0 failures

Shared, reproducible research state can let many coding agents compound each other's progress instead of repeatedly restarting the same search.

One useful ML email a week.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.