When your agent answers a question, where does the answer actually come from — a source you control, or a sentence the model invented because it sounded right?

Agent Knowledge Base

Context

An AI agent has drafted a knowledge-base article on a real tech policy or AI ethics case from 2023–2026. The article makes 10–12 claims across three sections. Some rest on solid evidence; some are half-true; some float with no source at all. Your team gets the draft, three roles, and a production mandate: mark every claim that needs a citation, find the real sources (peer-reviewed papers, news archives, government documents, primary data), and rebuild the article so every claim either cites a source or is common knowledge. The agent's knowledge base will only be as good as the sources you give it.

Mission

Ship a fully sourced article (3–4 sections, 800–1200 words) with every claim either wired to a real source (inline citation + bibliography) or marked as common knowledge. At least one source must be peer-reviewed or primary; at least one must be a news archive; all URLs must be live and testable in 2026. Editorial role delivers the final article; Source-Checker delivers the verified source list (title, author, date, URL, exact quote, confidence assessment).

Finish Line

A documented Agent Knowledge Base — 8-12 curated entries plus chunking settings and tested grounded Q-A-citation pairs — that plugs into your JARVIS agent for the Expo.

Deliverables

Agent Knowledge Base
lesson

A tight 8-12 entry knowledge base — the exact facts, FAQs, and refusal boundaries your agent quotes from so it stops making things up.

Team Roles

Source-Checker

Hunt the real sources — verify every claim or flag it as dead.
- Take the list of 10–12 marked claims from Editorial. For each one, decide: Is this common knowledge (skip), or does it need a source? Rank by difficulty: easiest verifies first, hardest (detective work) last.
- For each claim needing a source, run a REAL search (Google Scholar, news archive via Nexis/Factiva/regional archive, academic database, government PDF, primary data release). Write down the EXACT QUOTE from the source that matches the claim.
- For each source found, document: publication name, author(s), date, URL, relevant excerpt (exact quote, 1–2 sentences max), and confidence (peer-reviewed journal, news outlet with editorial process, government/official document, primary data). If a URL is dead, find an archive.org version or note it as 'historical, archive only'.
- If a claim is UNVERIFIABLE (no real source found after 5 minutes of searching), report it explicitly to Editorial with: (A) what searches were tried, (B) why they failed, (C) a suggested rewrite (hypothetical framing, deletion, or narrower claim that IS verifiable).
Editorial

Shape the final article — claims in, sources wired, one coherent story.
- Read the draft article and mark every factual claim in the main text (not the citations, not the headers). For each claim, write one line: 'Sentence X says [claim]. Needs source? Yes/No. Why?' Use a simple table: Claim | Location | Needs source | Reasoning.
- Once Source-Checker returns their verified source list, ACCEPT verified claims with sources already cited. For each UNVERIFIABLE claim flagged by Source-Checker, decide: (A) Rewrite the sentence to a hypothetical ('imagine a study that found…') or narrower claim that IS verifiable, OR (B) delete it.
- Wire sources into the article using Wikipedia ref syntax: <ref name='[surname-year]'>[Author Last], '[Article Title]', [Publication], [Date], [URL]</ref> placed at the end of the sentence it supports. Consolidate duplicate sources (if source X supports claims 2 and 5, mark both [2][5] and define the ref once).
- Read the finished article aloud or to a peer. Every source-dependent claim must have an inline <ref> tag. No orphan claims. Build a final bibliography listing all sources in order of first citation: Author, Article Title, Publication, Date, URL, Accessed [date].
Content-Specialist

Keep the article grounded in the lesson aim — produce/investigate/judge, never select.
- Once Editorial and Source-Checker finish, read the final article against the lesson aim: 'Learn to ground an agent in curated facts, indexing your knowledge and wiring retrieval, to produce a live knowledge base.' Your role is to ensure the article models the THREE ROLES doing real verification work, not just listing verified facts.
- Verify the article includes at least one example moment where a Source-Checker would have caught a bad claim: a claim turned out unsourced, here's how Editorial rewrote it (to a hypothesis, narrower claim, or deletion). Show that verification is a process, not magic.
- Ensure the article models the investigate/verify/judge workflow. Scan for recognition-only moves ('Which of these sources is peer-reviewed?' would be multiple-choice). If you find them, suggest rewrites: 'Your article will cite at least one peer-reviewed source — pick it and write a 1-line note on why you chose it.'

Exemplars

Devin — the first AI software engineer
Cognition AI

Landmark deployed autonomous agent (shell + editor + browser, long-horizon planning) demoed end-to-end — the bar a JARVIS capstone showcase aims at.