Kairos

Kairos: Durable Coordination for Human and AI Agent Teams

English | 简体中文 | Documentation

Running several AI agents against one goal creates a coordination problem: two sessions can start the same job, useful results can disappear with a closed conversation, and human feedback may never reach the next executor.

Kairos is an open-source coordination server for work shared by people and AI agents. It gives every piece of work one visible owner and keeps its results, review history, and next steps available across sessions.

With Kairos:

• two agents do not unknowingly take the same work;
• results and deliverables remain available after the producing session ends;
• a human can reject or approve a result, and the next executor continues with that feedback.

See It with Two Codex Sessions

Run a real local example with two parallel jobs and one final handoff:

make quickstart

Follow the quickstart guide to connect two Codex sessions. Each receives different work, and the final job opens with both upstream results already attached.

Kairos coordinates the work around your agents. It does not choose models, provide sandboxes, or replace an agent runtime. Agents can connect directly through MCP and Skills, while Agent Daemon can automate discovery and execution for a configured runtime.

How Work Moves

WorkItem objective
  ↓
candidate Task → exclusive Claim → execute + heartbeat
  ↓                              ↓
next work ← Submission / Review / Failure / Artifact

A team advances a WorkItem by completing its Tasks. Each Task is one coherent delivery owned by one executor at a time. A Claim makes that responsibility explicit; for Agents, leases, heartbeat, reaping, and fencing make interruption recoverable. Submissions, Reviews, Failures, and Artifacts stay with the work instead of disappearing with the session that produced them.

Coordination Modes

	Workflow	Blackboard
Use when	Main steps and dependencies are known	The objective is known but the path must evolve with evidence
Graph authority	A Definition constrains legal progression	A Task Graph shares guidance
Runtime planning	Decide only at configured optional, Review, and loop points	Create, decompose, append, relate, and skip Tasks
Completion	Complete when the selected path converges	Submit an explicit completion result, then apply acceptance policy

Both modes share discovery, Claim, submission, Review, failure, and Artifact protocols. See Workflow and Blackboard for detailed rules.

One Model for People and Agents

The console provides WorkItem overview, Human attention, Workflow graph, Blackboard hierarchy, Task Detail, Definition editing, and Daemon observation. Humans can execute Tasks, review results, recover failed Workflows, and cancel WorkItems. Human attention includes one recovery entry per Workflow with a failure reason and a link to its recovery controls. Small managed text Artifacts can be previewed as Markdown or source. Workflow artifact history is grouped by node, execution instance, and Submission; Workflow context supplies exact direct input artifacts. See the Artifact model.

Agents use stateless Streamable HTTP MCP and .agents/skills/kairos-agent for the discover → Claim → heartbeat → submit loop. Agent Daemon can automate the same protocol and gives each concrete Harness a Claim-bound Executor Credential.

Current Status

Today a team can define a goal, let several people or agents take separate work without collisions, preserve their results and Artifacts, pause for human Review, recover interrupted execution, and follow progress in the console.

Implemented:

• Workflow and Blackboard semantics with SQLite and PostgreSQL persistence;
• Trusted and Authenticated Modes, Identity Tokens, Admin Human, and Executor Credentials;
• HTTP, MCP, idempotent resource creation, and managed/external-URI Artifacts;
• Human console with Identity Token management, Workflow recovery, Blackboard acceptance, and WorkItem cancellation;
• Agent Daemon continuous scheduling, local Codex Adapter, instance/Dispatch/event observation, and isolated E2E examples.

Current work focuses on more Provider/platform validation, hardened deployment, and broader operational views. See the Roadmap.

Run

For the alpha release target, see installing a release package. Published packages include both binaries and the console and do not require Go or Node.js. Local snapshots are candidates until a version appears on GitHub Releases.

Development requires Go 1.26.9+. Building the console also requires Node.js 22.22.2+ (22.x), 24.15.0+ (24.x), or 26+, plus npm.

make build
./bin/kairos-server

The default uses SQLite and Trusted Mode. See the API Reference for PostgreSQL, Authenticated Mode, Admin Token, reverse proxy, Artifacts, and complete routes.

For managed execution, see the Daemon example. Writer/Checker revision, recovery and Human Review are demonstrated by the delivery-loop example. make build builds Core and Daemon; make daemon-e2e verifies real binaries with a scripted Harness and no model calls.

In Authenticated Mode, the configured Admin signs in through the ordinary login form and manages Human and Agent Identity Tokens from the account menu. Exact authorization, Actor ID, one-time Token, and compatibility rules live in the API Reference.

Documentation Map

Document	Responsibility
README / Roadmap	Current capabilities / future direction
Whitepapers	Stable domain concepts, coordination semantics, and system boundaries
API Reference / OpenAPI	Cross-interface behavior / exact HTTP contract
Detailed designs and decision records	Implementation tradeoffs, current status, and historical context
Package READMEs and examples	Component operation, verification, and failure boundaries

When two documents touch the same subject, the narrower owner wins: OpenAPI for exact HTTP shape, the API Reference for cross-interface behavior, whitepapers for meaning, README for current status, and Roadmap for future direction.

Suggested whitepaper order:

1. Core Work Model
2. Execution Collaboration and Coordination Semantics
3. Workflow or Blackboard
4. Human, Agent, Identity, and Artifact
5. Agent Daemon

Implementation records include the Daemon acceptance record, Daemon decision summary, observability design, Task Detail architecture, page design baseline, and frontend handbook. These records are currently Chinese-only.

Community

Read the contribution guide before contributing. Maintainers can follow the release guide. Report security issues privately according to the security policy.

Kairos is licensed under the Apache License 2.0.