Multi-agent collaboration
Optional — requires the collaboration module.
Got a spec with seven tasks, in waves, with dependencies, and want several agents to split the work? Enable the collaboration module (tick it in the wizard, or turn it on from Project settings with the project open), then:
1. Give tasks an assignee and dependencies
Section titled “1. Give tasks an assignee and dependencies”A task can carry an assignee and a set of prerequisite tasks. Until every dependency is done, the engine blocks both starting and completing the task — “don’t start a task whose dependencies aren’t finished” is guaranteed by the machine rather than by self-discipline.
2. Give each agent an identity
Section titled “2. Give each agent an identity”Set AIPOS_AGENT in each agent’s environment (fe, data, …), or add --agent to the launch arguments in .mcp.json. Every ledger entry then carries that identity, and the collaboration page’s activity feed shows who did what.
3. Claim vs. assign
Section titled “3. Claim vs. assign”An agent claims a task with claim_task — claims cannot be stolen; claiming an already-claimed task fails. You, or a dispatcher agent, use assign_task to reassign (this can override, and every handover is recorded).
These two tools only appear in projects with the collaboration module enabled — both in the Tools page authorisation directory and in the agent’s own tool list. A project created with the module already ticked has them authorised from day one; if you switched the module on later, check they are enabled on the Tools page (see Connect an AI agent).
4. Source code doesn’t collide
Section titled “4. Source code doesn’t collide”What AI-POS guarantees is the consistency of tasks and their state. If several agents edit code at the same time, give each one its own git worktree or branch and merge with git at the end. That boundary is written into the collaboration module’s own guide, which ships into the workspace when you enable the module.