Relic: collaboration lessons that outlive the agents who learned them
Sep 26 Du et al.: Relic turns recurring multi-agent friction into organization-owned executable protocols — complete-contract delivery 14.06%→19.76% over matched structured teams; fresh-member transfer 41.2% with bindings vs 34.6% text vs 25.4% none.
Source: Relic: From Multi-Agent Collaboration to Persistent Organizational Capability
Hongyi Du, Tianyi Zhang, Weijia Zhang, Yi Yang, Haofei Yu, Kunlun Zhu, Tianxiang Dai, Shang Jiang, Zhelun Gao, Jiaxin Pei, Shang Zhu, and Jiaxuan You start from a familiar factory failure: one coding agent changes an interface while another keeps building on the old version until tests go stale. A conversation can fix the episode; when the participants rotate, nothing remains that still governs the team. Relic turns recurring collaboration failures into organization-owned, executable protocols—members reflect on visible work, propose rules, and govern adoption. Adopted protocols bind triggers, responsibilities, required evidence, and runtime consequences, and stay open to revision and retirement.
In a traced case, repeated integration friction becomes an interface-review rule that later governs pull requests and is revised as work continues. Across 360 controlled runs over ten software workloads and three models, Relic lifts complete-contract delivery from 14.06% to 19.76% (+5.71 pp) over a matched structured team without the protocol lifecycle, improving all four verified production endpoints in every model stratum. Under fresh-member transfer, behavioral correctness is 25.4% with no inherited protocol, 34.6% when the same rules are readable text, and 41.2% with executable bindings—a +6.5-point advantage of bindings over text. On CooperBench (excluding broken pairs) Relic hits 367/477 (76.9%), best among peer-structured systems; on the fixed 48-pair same-model subset it beats Solo (29/48 vs 26/48), reversing the official peer baseline’s coordination loss.
Why it matters for code review
Overnight and always-on multi-agent factories accumulate coordination debt faster than chat memory can hold it. Treat learned integration rules as reviewable harness software: require protocol triggers and evidence gates in the runtime, not just AGENTS.md prose, and score whether a new agent cohort still obeys last week’s lessons. Continuous verified momentum compounds when the organization—not the session—owns the fix; FOMO for more concurrent agents without executable handoff protocols just recreates the same interface-stale failures every morning.