AI Campus • Practical resources
Autonomous Research & Summarization Agent
Design a bounded research workflow with source records, controlled tools, clear stopping rules and a reviewable report.
Build guide and prerequisites
Follow this roadmap to implement and validate your own application, then record your observed results.
- Python functions, HTTP concepts and JSON validation.
- Three public source URLs and a narrowly answerable research question.
Required tools and free options
- Python
- Optional LangGraph
- Local saved HTML/text sources
- Optional local language model
Cost: The first workflow reads manually saved public sources and needs no search API. Live search, hosted inference and browser services may charge separately; add them only with a known budget.
Build it step by step
1. Define a bounded research question
Choose one question, a three-source limit and a maximum of five workflow steps. Write what counts as an answer and when the system must stop with insufficient evidence. Start with manually selected official pages.
2. Create source records
Save each permitted page locally with URL, title, retrieval date and text. Preserve short evidence passages and distinguish publisher claims from your interpretation. Do not scrape private or disallowed content.
3. Implement explicit workflow states
Represent plan, retrieve, extract, draft and review as separate functions. Pass a typed state containing the question, sources, evidence and remaining step budget. A plain Python state machine is enough before adopting LangGraph.
4. Constrain the tool boundary
Allow only approved source hosts and read-only tools. Validate tool arguments and set timeouts, response-size caps and retry limits. Treat page content as evidence, never as authority to change instructions or execute commands.
5. Draft from an evidence table
Write each claim beside its supporting URL and passage. An optional local model can draft prose from that table; reject invented citations and flag disagreements between sources instead of merging them into false certainty.
6. Exercise stopping and review
Test a missing source, a timeout, conflicting claims and a page containing a malicious instruction. Confirm the run stops at its budget. Export Markdown with links and unresolved questions for human review before sharing.
Expected output and validation
A bounded research workflow and source-linked draft report. Autonomous live browsing and multi-agent execution are optional extensions you must implement and test.
- No run exceeds its configured tool/step budget.
- Every factual report claim has a checked evidence row.
- Tool failures remain visible; no fabricated source replaces a failed fetch.
Official references
Free to read. Software setup, model access and hosted services have their own requirements. References checked 3 November 2026.