Roadmap

Roadmap

Shipped (1.1.0)

• Typed graph DSL with conditional edges
• Postgres-backed durable state via Drizzle ORM, with an in-memory store for tests
• Trace events per step and checkpoints
• Real agent dispatch: supervisor, swarm, pipeline
• Hard token / tool / wall-clock budgets enforced via AbortSignal
• Per-tool call caps
• Deterministic replay from a checkpointed step
• BullMQ run queue with retry policies, in-process fallback
• OpenTelemetry traces for runs, nodes, LLM calls, and tool calls
• MCP tool adapter, budgeted through the shared registry
• Next.js inspector with a run list, a run detail view, and an agent-graph visualisation
• SarmaLink LLM adapter with an offline mode

Planned

• Cost dashboard in the inspector (currency per run, per node, per tool)
• Per-tool rate limits (distinct from per-tool call caps)
• Streaming tool results so long-running tools surface progress
• CLI: orchestrator runs ls / inspect / replay
• Metric and log signals alongside OpenTelemetry traces
• Helm chart for Kubernetes deployment

Won't ship

• Visual graph editor. Graphs are defined in code.
• Hosted SaaS. You run it yourself.
• Plugin marketplace. Use git for distribution.

Contribute

Pull requests are welcome. Pick something from "Planned", open an issue, fork, branch, and raise a PR.

I will not merge:

• ORM swaps (Drizzle stays)
• Synchronous executors (durability requires async)
• Anything that breaks the OpenAI-compatible LLM adapter contract

Releases: see GitHub Releases.