Replace the agent.
Keep the work.
Agents stop, fail or hand work to other agents. Solu keeps saved results and permissions with the work, so changing agents does not mean starting over.
- Keep the work in one persistent place as agents come and go.
- Give agents time-limited responsibility for the work. An agent can pass on only permissions it already has.
- Keep saved results available when an agent is replaced.
- After a restart, require agents to prove they still have permission before continuing.
- Let the service receiving a request check that the agent is allowed to make it.
Work
What is the task, and what counts as accepted?
Control
What stays live, and what must stop?
Continuity
What survives when an agent changes?
Observation
What actually ran?
Team memory
What is happening across the teams?
One coordinator. A durable record.
Solu is built in Rust. Every transition is validated, appended to a checksummed write-ahead log and synced to disk before it becomes live. Recovery replays the same transitions. Retained result bytes are content-addressed and verified on every read.
Operations hold leased custody. Grants suspend when custody is lost and expire when it ends. Providers verify short-lived Ed25519 proofs offline; an already-issued proof remains valid until expiry, up to 30 seconds.
Current scope: a working prototype with one coordinator per instance. There is no distributed consensus, sandboxing or automatic log repair. Remote agent traffic is signed but not encrypted.
- Built with
- Rust, a durable write-ahead log and Ed25519 proofs.
- Retains
- Instance identity, custody history, grants and retained results.
- On failure
- Custody becomes lost; recovery requires proof. A corrupt log tail needs an operator.
- Boundary
- Preserves the work; does not isolate or schedule agents.