Multi-step research with a real state graph: planner → search × N → distill → verify → write. Every run is a replayable trace.
Periscope is a research agent that runs ten-step research workflows over heterogeneous sources — arXiv, Scholar, the open web (via Tavily), code repositories — and produces a sectioned report whose every claim is attributed back to a retrieved chunk.
Every run is exposed as a replayable trace with span waterfalls, per-step tokens/latency/cost, and the exact prompt/response at every node. Cron schedules and saved searches turn the agent into a continuous research surface, not a one-off chatbot.
- Real state machine, not a chain — LangGraph nodes for
planner,search.{web,arxiv,code},fetch.pdf,planner.replan,synthesize,verifier.cite,write.report. Conditional edges loop back on insufficient coverage. - Durable checkpointing — every step persisted via
langgraph-checkpoint-postgres. Process crashes mid-run, restart, resume from the last checkpoint. - Traced end-to-end — LangSmith integration plus a built-in trace inspector. Span waterfalls show tokens / latency / cost per node; the entire prompt/response history is retained.
- Citation-grounded reports — Pydantic-validated structured output; a citation-coverage verifier independently scores every claim against retrieved chunks (average grounding 0.96).
- Operable — cron schedules (hourly arXiv watch, weekly RAG digest), saved searches, fork-a-run with a different model.
 |
 |
 |
 |
| Layer | Tech |
|---|---|
| Agent | LangGraph 0.2.50, langgraph-checkpoint-postgres, LangChain 0.3, langchain-anthropic |
| Models | Anthropic Claude sonnet-4-6 |
| Search | Tavily search API, arXiv API |
| HTML clean | BeautifulSoup, readability-lxml, lxml |
| Persistence | Postgres 16 (checkpointer), psycopg[binary,pool] |
| Observability | LangSmith, structlog, custom trace inspector |
| Ops | Tenacity (retries + circuit breakers), Pydantic 2, Streamlit operator console, Docker Compose |
git clone https://github.com/phantomdev0826/periscope-agent
cd periscope-agent
cp .env.example .env # add ANTHROPIC_API_KEY + TAVILY_API_KEY + LANGSMITH_API_KEY
docker compose up -d --build
docker compose exec agent alembic upgrade head
docker compose exec agent python -m agent.cli run \
--question "What are the practical tradeoffs of speculative decoding?"A typical run lands in 38–51 seconds, 9 steps, 6 sources, ≈ 14k tokens / $0.08. Open the trace inspector in the Streamlit operator console (http://localhost:8501) to step through it span by span.
┌─────────┐
│ planner │
└────┬────┘
┌───────────┼───────────┐
┌──────▼─────┐ ┌───▼────────┐ ┌▼─────────────┐
│ search.web │ │ search. │ │ search.code │
│ (Tavily) │ │ arxiv │ │ (GitHub) │
└──────┬─────┘ └───┬────────┘ └┬─────────────┘
└───────────┼───────────┘
│
┌──────▼──────┐
│ fetch.pdf │
└──────┬──────┘
│
┌──────▼──────────┐
│ planner.replan │ ──── coverage gap → search again
└──────┬──────────┘
│
┌──────▼──────┐
│ synthesize │
└──────┬──────┘
│
┌──────▼─────────┐
│ verifier.cite │ ──── grounding < 0.7 → revise
└──────┬─────────┘
│
┌──────▼───────┐
│ write.report │
└──────────────┘
▲ every node checkpointed to Postgres — resumable, replayable, forkable
docker compose exec agent pytestIncludes tests for graph wiring, the citation verifier, the planner's replan logic, and Pydantic schema enforcement on the final report.
MIT