ADR 0009: Peer-to-Peer Architecture

Date: 2026-06-27 Status: Accepted

Context

We want to test a linear handoff model where two peer agents share context via a message bus, without a central supervisor. The key behavioral question is: can a sequential peer pipeline transfer enough structured context between agents to complete a compound task without a coordinator?

Decision

Implement a PeerToPeerArchitecture with: - CRM Peer (Agent A): receives the original task, performs all CRM lookups, and produces a HANDOFF: structured message as its final output. - Ticketing Peer (Agent B): receives the original task prompt plus Agent A's handoff message concatenated as its prompt. Completes all ticketing operations using the extracted context.

The shared "bus" is a simple Python string: Agent A's output is appended to Agent B's input message. No LLM is involved in the handoff itself.

A peer_handoff trace event is emitted between the two agents, recording the handoff message, enabling the grader to verify whether context was successfully transferred.

Trace Events

Adds: peer_handoff, run_finish.

Consequences

• Positive: Simplest multi-agent topology. No coordinator overhead. Handoff is transparent and fully traceable.
• Negative: Agent B is entirely dependent on Agent A producing a well-structured handoff. If Agent A fails (tool errors, wrong format), Agent B has no fallback and no ability to re-query Agent A. The architecture is vulnerable to first-hop failures.