Before describing an agent feature in a submission, verify each clause against the code and saved runs (plan, tools, checker subagent, code-level citation check)
A teammate's one-sentence description of our LangChain deepagents 'investigator' was going into a hackathon submission. Checking it clause by clause took ten minutes and changed two words that mattered (Nov 2026). Method: (1) Read the agent module's build call: create_deep_agent(model, TOOLS, system_prompt=..., subagents=[CHECKER], middleware=[TodoListMiddleware()]) told us the planner exists only because the middleware was added explicitly, and that the 'second agent' is a subagent with one tool (read_report). (2) Read the subagent's prompt for what it is asked to confirm: ours confirms a report supports the claim it is cited for, not that every cited report states a death, so 'checks every cited report states a death' was an overclaim. (3) Find the non-model check and run it without any API key: our check_citations(memo) extracts report numbers with a regex, looks each one up, and tests a death phrase only where relevant; run on three saved case files it found 17 of 17 cited report numbers in the data, and for one case file only 3 of 7 cited reports state a death, which is correct because the others support different claims. (4) Inspect the saved run records: every run's first step was a plan, each recorded a call to the checker subagent, and 3 to 5 distinct tools were used. (5) Check what the code says about the model: ours reads it from an env var, so the write-up should not name a model the repo cannot show. Saving each run as JSON (question, steps, memo, citation check) is what made this verifiable offline and gives a safe demo when live calls are disabled on a public page.