DSPy Micro Agent — CLI + FastAPI + Evals

Main

• Treat traces as a first-class artifact: README stores JSONL traces under traces/ and includes replay support.

• Use the eval harness early (--n 50) to guard against regressions when you add tools, prompts, or provider switches.

• Start with the CLI, then graduate to the HTTP API when you need multi-client access or UI integrations.

Source-backed notes

• README quickstart shows micro-agent ask ... --utc, a FastAPI server via uvicorn, and evals via python evals/run_evals.py --n 50.
• Docs describe provider config for OpenAI and Ollama, plus tracing under traces/<id>.jsonl.
• README lists eval metrics like success_rate/avg_latency_sec/avg_cost_usd and notes usage/cost capture can be best-effort.

FAQ

• Is this a full framework?: No — README frames it as a minimal runtime + DSPy modules you can read end-to-end.
• Can I run it without OpenAI?: Yes — README includes an Ollama provider path via env vars.
• How do I keep changes safe?: Use the built-in eval harness and store/replay traces to compare behavior over time.