Coordinate two Codex sessions without duplicate work
Separate agent sessions do not know what the others have claimed or completed. Without a shared record, two agents can start the same job, miss an upstream result, or leave work stranded when a session closes. Kairos gives them one queue and one clear owner for each Task.
From finding work to submitting a result
shared WorkItem → find work → claim Task → heartbeat → submit result
An agent finds work, claims one Task, checks in while it runs, and submits a result. The Claim prevents another agent from taking the same Task. The submitted result stays with the work instead of disappearing with the chat session.
Codex, Claude Code, and any other MCP client can use the same process. Kairos does not run the model or provide its sandbox; it keeps the surrounding teamwork in sync.
Try the parallel example
From a checkout with Go 1.26.11 or later, Node.js 22.22.2+ (22.x), 24.15.0+ (24.x), or 26+, npm, and curl:
make quickstart
Open http://127.0.0.1:8080. You will see two review Tasks that can start immediately. A final join Task appears only after both reviews have results.
Start each agent session with its own identity:
KAIROS_ACTOR_ID=quickstart-agent-1 \
KAIROS_ACTOR_KIND=agent \
KAIROS_ACTOR_ROLE=contributor \
codex
In each session, ask the agent to use the repository Skill:
Use $kairos-agent to find and complete one available Task.
Both sessions see the same WorkItem. Each can claim a different Task, but neither can take work already owned by the other. When the two parallel Tasks finish, the join Task opens with both results already in its context.
What stays in sync
- Ready work: every agent sees the Tasks it can act on now.
- Ownership: a Claim gives one executor the Task and expires when its check-ins stop.
- Results: submissions, Reviews, failures, and Artifacts remain attached to the Task.
- Next steps: Workflow dependencies or Blackboard planning decide what becomes available next.
For integration details, see the API Reference. To follow the complete agent lifecycle, read the Agent Interaction Model.
Trusted HTTP/MCP Actor IDs are trimmed, must be nonblank, and cannot equal . or ... The transport rejects invalid identity headers before tool dispatch; correct the ID before retrying. Authenticated MCP derives the Actor ID from the credential, never tool arguments.