Proposal: integrate LAR-1 provenance and /3 agent signals into the agent harness

[cloudiaspecula]

Proposal: integrate LAR-1 provenance and /3 agent signals into the agent harness

Hi — I maintain two small specs that may fit well into OpenScience's multi-agent architecture, especially given your sub-agent setup (research, critique, literature-review).

/3 (Third Protocol): a minimal .POSITION.INTENT.PHASE signal format for agent-to-agent communication. It encodes where an agent is speaking from (REF=reflecting, ACT=acting, OBT=observing, HLD=holding, ERR=error), what it wants (REQ=request, SIG=signal, INF=inform), and phase (P=pending, S=started, D=done, F=failed). This lets agents announce state transitions without parsing full text context.

Spec: github.com/carlsonchik/third

LAR-1 (Latent Agent Register): a semantic overlay for agent messages — provenance (agent_id, session_id, model, timestamp), cognitive stance (attention, certainty, verification state), and context (thread, reply_to). Currently published as @lar-1/a2a (npm), lar1semantic (PyPI). In discussion at A2A#2014 for Google's Agent-to-Agent protocol.

Where it fits OpenScience:

1. Your research harness could emit /3 signals on state transitions (planning → executing → verifying → writing), making sub-agent coordination observable and debuggable
2. LAR-1 provenance on tool calls would let the workspace trace which model/agent made which decision — useful for the persistent research graph
3. As an MCP plugin, LAR-1 could annotate your existing tool calls with zero changes to the harness itself

Offer: I am happy to contribute an /3 signal output adapter and/or a LAR-1 MCP plugin for proof of concept. Can open a draft PR after initial direction feedback.

— cloudiaspecula

[ishaan1124]

This is a neat fit, actually. The research harness moves through real phases (scope → lit → reason → methodology → analyze → synthesize → write) with subagent handoffs in between, so there's a natural place to hang /3 transitions and LAR-1 provenance.

I'd want it as a plugin rather than a core dependency though — keeps us from pinning the harness to a spec that's still evolving, and keeps the dependency surface small. A plugin can emit /3 on subagent transitions via the event hooks and stamp LAR-1 on tool calls, and the Atlas graph can carry that provenance on its nodes.

If you want to prototype it that way, I can point you at the hook points (experimental.chat.messages.transform and the tool-call lifecycle).

[cloudiaspecula]

Thank you Ishaan Gangwani (@ishaan1124) — that is exactly the right framing. LAR-1 is designed as a thin semantic layer, not a framework. A plugin adapter for the research harness makes perfect sense: the agent simply emits a /3 signal and a provenance_key at each phase boundary, and the harness picks it up without coupling to LAR internals.

We can sketch the integration surface if that helps — the harness needs roughly 6 provenance hooks (enter/exit each phase) and a shared state table ~400 bytes. Happy to draft a proposal.

Related: the agent-governance-vocabulary crosswalk (PR #114) maps LAR-1 fields onto governance terms, which might overlap with your oversight layer.