Roadmap
v0.1 — now. Provider layer (Anthropic + every OpenAI-compatible host), @tool, the ReAct loop with real streaming and parallel tool calls, six tool packs, MCP client with registry, handoffs, orchestration (worker.as_tool() — a supervisor agent that delegates and keeps control), multi-turn via history=, CLI. Project scaffolding: nobelagentic new my-agent (Python) and npm create nobelagentic (TypeScript). Python and TypeScript with the same API and the same event stream.
v0.2 — the production agent. The graph kernel: nodes, edges, typed state, checkpoints and resume. Agent is re-based onto it so human-in-the-loop approval ("pause before this tool runs, ask a human") and durable runs come for free. Plus the three things every production user asks for first:
- Structured outputs —
Agent(..., output=MyModel)returns a validated Pydantic model / zod object, not a string. - Sessions —
agent.run(msg, session="user-123")with pluggable storage. Multi-turn without hand-rolling history. - Run limits —
max_turns, token and cost budgets per run, tool timeouts, provider retries with backoff. - Native Gemini adapter.
v0.3 — crews, memory, and eyes. Crew(agents, tasks, process). Short-term and vector memory. Knowledge sources — point an agent at your docs (knowledge=["./docs/*.pdf"]) and retrieval is handled. Input and output guardrails. Multimodal input (images and files straight into run()). Best-in-class observability: OpenTelemetry tracing, per-run cost tracking, and a local trace viewer.
v0.4 — testable and everywhere. Evals: record a real run once, replay it offline forever — assert on tool calls, outputs and cost in CI. A sandboxed code executor pack (the safe version of shell). MCP server mode (nobelagentic serve --mcp) so your agents are tools for any MCP client. A tool registry: nobelagentic add tool/weather pulls a community tool the way npm pulls a package. 100+ built-in tools, many as thin wrappers over maintained MCP servers.
v0.5 — developer experience. nobelagentic dev — a local playground over the event stream: chat with your agents in the browser, watch tool calls, handoffs and token spend live. nobelagentic serve — any agent as an HTTP endpoint with SSE streaming. Lifecycle hooks (on_tool_start, on_agent_end, …) and a typed run context, so tools receive your app's state (a DB handle, the current user) instead of reaching for globals. A documented wire protocol and React hooks for streaming agent state straight into your UI.
Non-goals, for now. Voice/realtime pipelines and hosted inference. Plenty of good products do these; we'd rather be the best framework than a mediocre platform.
Suggest or vote on features in GitHub issues.