中EN

Multi-agent systems · Organizational learning

RelicFrom Multi-Agent Collaboration to Persistent Organizational Capability

  • Hongyi Du1*†‡
  • Tianyi Zhang2*
  • Weijia Zhang3*
  • Yi Yang1
  • Haofei Yu1
  • Kunlun Zhu1
  • Tianxiang Dai4
  • Shang Jiang5
  • Zhelun Gao6
  • Jiaxin Pei7
  • Shang Zhu8‡
  • Jiaxuan You1‡

1 University of Illinois Urbana-Champaign2 Harvey Mudd College3 Yale University4 Stanford University5 Peking University6 National University of Singapore7 University of Texas at Austin8 Together AI

* Joint first authors; author order reflects relative contribution. † Team lead. ‡ Corresponding authors.

Turning recurring collaboration friction into organization-owned, governed, executable protocols that persist across members.

Preprint · arXiv:2609.32965

How protocols form, execute, and persist across members

Open figure
Click or hover to follow how a protocol forms, guides execution, and persists as members change.

The idea

What does a team keep when its members change?

Sharing a repository does not prevent conflicting changes, stale tests, or recurring coordination failures. When members change, what keeps the team’s agreements working?

Relic turns collaboration experience into organization-owned, executable protocols. Members propose, review, and adopt rules that specify triggers, responsibilities, and required evidence. These protocols persist beyond individual members and shape later work, while remaining open to revision and retirement.

Read the full paper

Key results

From shared experience to measurable capability

Follow each result to its comparison and full estimates
Main experiment+5.71 pp

Complete mainline contracts

19.76% with Relic · 14.06% with B2

360 runs across three models; paired gain over B2, 95% CI [2.97, 8.55] pp.

Evidence in the paperInternal transfer+15.8 pp

Behavioral-case pass rate

Exec 41.2% · Fresh 25.4%

Fresh members with executable protocols versus no inherited rules; 95% CI [9.2, 31.8] pp.

Evidence in the paperCooperBench79.1%

Full-benchmark successful pairs

Relic 371/469 · Peer 59.1% · Team 74.4%

All 652 benchmark pairs were audited; the same 183 broken-pair exclusions leave 469 final-valid pairs for every system. Separate 47-pair same-model check: Relic 28, Solo 26, Peer 13.

Evidence in the paperProtocol formation280

Formed protocol lineages

Sustained use across 60 Terra + Opus Relic runs

32 also link use by other members to shared-state changes and evaluated outcomes.

Evidence in the paper

Explore further

Mechanism, evidence, and a working organization

01

Evaluating organizational capability

The main study, runtime-binding ablation, and member-transfer experiments examine how protocols affect work.

Explore the evaluation02

Take the human seat

Ask the secretary for a work briefing, examine a decision, and follow changes across members, code, and the organization.

Open the experience03

Build on the research

Reproduce the paper with Relic, run your own organization with Relic-Agent, and explore protocols and governance.

Browse resources