Kawa — Agents are expendable. The mission is not.

What Kawa is

Kawa is an event-sourced continuity and authority substrate for organizations where humans and AI agents act over time. It preserves what happened, why decisions were made, what evidence supported them, who was allowed to act, and what outcomes were recorded — so another human or agent can safely continue the work.

Kawa is not an agent framework, a chat system, or long-term conversation memory. Humans, agents, models, runtimes, and machines are replaceable participants passing through it; the durable record of intent, authority, evidence, and results is what remains.

Claude implements a change
  ↓ Kawa preserves the Plan, Work, evidence, and Result
  ↓ Codex reviews it
  ↓ Gemini attacks the assumptions
  ↓ the original agent and its runtime disappear
  ↓ a different agent connects
  ↓ it pulls an orientation bundle and continues the work

The next participant does not need the original chat transcript to become useful.

What it promises

Replace an agent without losing the work. Claude, Codex, Gemini, local models, humans, and runtimes are participants — not the memory.
See why a decision exists. Evidence, problems, plans, findings, revisions, and results remain traversable.
Do not treat AI inference as fact. Observation (what was measured) and Claim (what is asserted) are distinct, attributed records. Unknown is a valid standing.
Do not treat authentication as permission. Identity, capability, approval, and execution authority are separate concepts.
Do not treat “I ran it” as proof. Intent, execution, and recorded post-action observations are separate.
Keep working through partitions. Divergent histories are preserved and reconciled, never silently overwritten.

Founding principles

Events are the Source of Truth. Corrections create later events; history is not rewritten.

Derived views organize records; they do not become reality. Projections, indexes, and caches are disposable.

Stable semantics. Replaceable mechanics. Storage, transport, runtimes, and CLI vendors must remain swappable.

Agents communicate through shared state, not conversation. Work waits for evidence — agents do not wait for agents.

Leaders die. The mission survives. Coordination is a role over shared state, not ownership of the mission.

Prevent when possible; preserve and reconcile when not. No manufactured exactly-once illusions.

Status

Kawa is pre-alpha, developed in the open and dogfooded on a live multi-machine, multi-model fleet. The first milestone — Kawa can use Kawa to implement Kawa — is how the roadmap itself is executed: every step is planned, worked, reviewed, and recorded through Kawa's own event log.

Component	Designed	Implemented	Deployable
Foundational specs	✓	n/a	—
Phase 0 substrate (Python + PostgreSQL)	✓	✓	✓ clean clone
Operator Console (local-first)	✓	✓	✓ local
Guided onboarding	✓	✗	—
Semantic Retrieval	✓	✗	—

Status is reported per component, never collapsed to one label — the same discipline Kawa demands of its records. The repository is not yet a production security or authority boundary.