Showcase 2 · Plan-and-Execute Autonomous Agent

Deep Research Agent

Submit a topic — the agent decomposes it into sub-questions, searches sources, cross-references findings, and emits a downloadable PDF. Every thought streams live so you can see the reasoning unfold.

1. plan Decompose topic
2. search ArXiv + web
3. factcheck Cross-reference
4. summarize Draft report
5. pdf Generate PDF
Try:
Live thought stream · idle

Submit a topic to begin. The agent will plan, search, fact-check, summarize, and emit a PDF report — streaming each thought live.

Time elapsed

0.0s

Time saved vs analyst

4h 0m

baseline: 4h manual research

Cost per report

$0.000

0 tokens

Cost saved

$200.00

vs $50/hr analyst

Sources verified

0

How it works

Streaming protocol

Server-Sent Events stream the agent's thoughts from the Astro API route to the React island as the pipeline runs. Each node — plan, search, fact-check, summarize — yields events that render directly into the live log.

PDF generation

The final report is rendered server-side with pdf-lib — pure JavaScript, no external services. Includes the executive summary, sources, citations, and run metadata.

Cost & time tracking

Per-node token counts are tallied against published API rates. The dashboard contrasts agent cost against the equivalent analyst rate (4-hour baseline, $50/hr).

Mock vs. real mode

Currently demos with deterministic mock sources. Real-mode swap-points are marked in the source: Tavily for search, Claude Sonnet for synthesis, OpenAI for fact-check grading.