Roadmap
Millstream is pre-release. The platform's five milestones are built, each closed by an end-to-end scenario that runs offline and deterministically. The open-source release comes next.
Milestones
| Milestone | Closed by | Status | |
|---|---|---|---|
| M1 | The core engine: pipeline file, packages, CEL, the SQLite store, deliveries, retries, dead letters, replay, the Go SDK | The reference pipeline's shape running end to end on a laptop, and the crash suite at every unit boundary | Built |
| M2 | Agents, containers, budgets and evals: filter, dedupe, the agent adapter, sandbox, transcripts, lessons, insights |
Scoring with an agent, a seven-day dedupe replay, and a lessons edit changing the scores | Built |
| M3 | Human review and the web UI, in English and French | The reference demo in the browser, both languages, no accessibility violation | Built |
| M4 | Server deployment: Postgres, standby and failover, remote nodes, shared streams, webhooks, retention and purge, Compose and systemd | The failover and node chaos suites, and a soak under random kills | Built |
| M5 | Self-improvement: fallback and audits, insight triggers, evidence, proposals, the improve package, distillation |
Production evidence yielding an accepted lessons and an accepted distill proposal |
Built |
| M6 | The open-source release: published binaries and image, signatures, the measured throughput ceiling, the documentation site | Next | |
| M7 | The commercial edition: single sign-on, roles, the first catalog actions | The first commercial deployment | After M6 |
After v1
Self-improvement
- Low-risk proposal classes applied behind a policy
- Lesson examples chosen per item by similarity
- Evidence pulled straight from production
Actions and agents
- A Python SDK
- Distillation into languages beyond Go
- A tool loop and MCP servers for the
apiharness - The Claude Agent SDK as a harness
Platform
- Network destinations enforced per container and agent
- Kubernetes nodes that launch action pods, and a Helm chart
- Vault and Kubernetes secret sources
Deliberately out of scope for v1
- A graphical pipeline editor. The pipeline file in git is the source of truth.
- A multi-tenant hosted service. Each customer gets a deployment of their own.
- Exactly-once side effects on external systems. Millstream passes an idempotency key; the external system enforces it.
- Sub-second stream processing. Agent steps need far less, and v1 publishes the ceiling it measures.
- Several active engines at once. One active engine with a hot standby.
- Kubernetes as a runtime. Nodes run on hosts in v1; pods and a Helm chart come after.
- Proposals applied without a person.