Agentic search
Agents have been excellent at finding pages and unreliable at keeping them. A subagent found a page, sent a summary, and the summary became the source. An image that held the fact could be missed. Will's Perimeter Solutions pieces needed every quote checked again. The fix is to save the page first, the way we save a filing, and let an agent read the file.
1. What a terminal skill does
- A link can be wrong, and a page can be skipped.
- An image that holds the fact can be missed.
- You cannot trust the summary as the page.
2. The script
- Company History asks three questions: founding, ownership, and milestones.
- Firecrawl saves the pages we open. We do not open every URL we find.
3. The agent sits on the files
- Unit economics is the same script with a new question and a new schema. Gross margin, cost per unit, revenue drivers. Fields: metric, period, quote, URL. The pattern, not a Costco result.
- Company industry is the same swap: a new question, a new schema, the same saved files.
- A company list is a later pass over sources we have already saved. It is not a trust list we build first.
Numbers read 3 October 2026 from ~/Code/company-history-capture/captures/costco/20261003T093324/manifest.jsonl. Allan's note is in InvestOS plan / priorities.