LOGOS: A Living Logic for AI Agent Teams That Evolve With Humans

What it adds beyond the summary

The summary states the architecture; the page supplies the governing principle, the formal objects, and the failure mode the design targets. The organizing rule is "proposal is not promotion": agents may propose changes, but release depends on independently defined evidence, policy, and authority — "a model should not both propose and approve its own updates." The named failure it exists to prevent is misevolution: a release artifact changes without evidence that the change improves the deployment (e.g., an ungated loop rewrites a prompt to raise a narrow proxy score when the actual defect is a changed tool schema, so the system changes while deployment utility falls).

How it works

The key distinction the summary omits is that the multi-agent system (MAS, denoted Π) is not the safety boundary. The governed deployment object is D = (Pack, R_root, b, Cred, S_env, Λ_audit), where R_root is a human-owned root policy sitting outside the ordinary self-evolution loop. It fixes objectives, evaluators, gate versions, protected holdouts, credential boundaries, permission-expansion rules, approval requirements, audit sinks, and allowed external effects. A candidate is a Pack′ evaluated under the fixed envelope (R_root, Cred, Λ_audit) against a sandboxed state copy — it cannot mutate credentials, live state, evaluators, or the audit sink.

The lifecycle is D --run--> (ζ, χ_eff) --propose--> D′ --gate--> D+. Two facts the notation separates: a proposal is not a release, and rolling back a pack is not the same as undoing an email, payment, or database write — external effects (χ_eff) require separate safeguards (previews, transactions, idempotency, compensation, human confirmation).

Portability across frameworks (it names AutoGen, MetaGPT, OpenHands, OpenAI Agents) rests on two contracts: a normalizer ϕ_b that maps each backend trace into a common trace space, and an emitter ε_b that re-expresses a portable lesson as a backend-native artifact — explicitly not a bit-exact replay, but a contract letting a skill learned on one backend be re-verified on another. Imported knowledge is re-gated on the target.

The five contributions: (1) a multimodal agent-team compiler emitting a probe-validated or fail-closed Agent Pack; (2) a pluggable execution/evidence layer with adapters, normalized traces, routing, and target-side re-verification; (3) an evidence-governed self-evolution protocol; (4) an auditable human–agent control contract (async direction, agent-initiated questions, scoped authority, approval, rollback, audit); (5) a bounded evaluation account separating task performance, mechanism evidence, fault injection, and contract conformance.

Limits

The fetched text is the abstract, introduction, and a primer through Section 2.2 — no benchmark results, numbers, or baselines appear in it. The paper claims evaluation via "external benchmarks, mechanism studies, generated fault injection, and contract-conformance checks," but the fetched portion shows the framework and formalism only, not outcomes. "Validated" is scoped narrowly: a ProbeValidatedPack is checked only against stated finite probes, not proven correct. The authors stress hard checks are "necessary but not sufficient." Anyone adopting this is adopting a governance contract and lifecycle, not a benchmarked system — the operational claims need verification against the full paper's Section 4.