Deep Research Agent
A plan-and-execute autonomous research agent — decomposes a topic, searches sources, cross-references findings, and emits a downloadable PDF. Every node streams its thoughts over SSE so the live trace is the demo.
A “Market Intelligence” agent that handles long-running autonomous research tasks. The user types a topic — for example, “NVIDIA H200 vs H100 benchmarks” — and the agent plans the work, runs it, and packages a PDF.
The pipeline
plan → search(×N) → factcheck → summarize → cost
Implemented as an async generator that yields thought events at every step. The Astro API route consumes the generator and pipes each event to the browser over Server-Sent Events.
Live thought stream
As the agent works, every internal step appears in the UI:
plan Decomposing topic into 4 sub-questions…
search Querying source A for "H200 inference benchmarks"
observe Found 3 candidate sources
factcheck Source A and Source B disagree on FLOPS — cross-referencing
summarize Drafting executive summary…
cost Tokens: 8,412 (≈ $0.07) · Duration: 11.3sThis turns the waiting time into a demonstration of agentic reasoning.
Tech choices
- Astro API route with an async generator — no separate Python backend, no WebSocket plumbing. SSE is enough.
- Vercel AI SDK (
ai) for LLM calls. Synthesis uses Claude Sonnet 4.5; the cheaper steps (planning, search drafting, fact-check grading) use Claude Haiku 4.5. pdf-libgenerates the final report server-side — pure JavaScript, no external services. Cover page, executive summary, sources, citations, and run metadata.- Cost tracking — per-node token counts are tallied against published API rates and shown on the completion card.
Mock mode
Currently demos with deterministic mock sources so the SSE protocol and PDF pipeline can be exercised without external dependencies. Real-mode swap-points are marked in the source for the search backend.