Multi-Agent Coordination Model

Multi-agent is a scheduling problem, not “open more chats”.

Coordination primitives

Tentarc coordination is built from:

• isolated instance keys
• explicit handoff events
• shared workspace-scoped context
• scheduler-level concurrency control

Recommended topology

For complex jobs, use role-split lanes:

• planner lane: decompose tasks and constraints
• executor lane: tool-heavy implementation
• reviewer lane: validation and risk checks

Each lane has its own instance stream; coordination happens via explicit events.

Handoff contract

A safe handoff event should carry:

{
  "from": "planner",
  "to": "executor",
  "taskId": "task_301",
  "inputRevision": 18,
  "expectedOutput": "patch + test result",
  "constraints": ["no schema migration", "no prod write"]
}

This avoids ambiguous “please continue” transitions.

Anti-patterns

• sharing mutable global scratch state between lanes
• implicit handoff via prompt text only
• mixing approval and execution ownership in same lane

Those patterns create race conditions and accountability gaps.