← back
Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers — Alex Bauer, Upside.tech
Takeaway
Build agent trust through clear intent, consistent business definitions, and shared guidance before expanding autonomous work.
Summary
- Agentic tools let go-to-market teams build solutions directly, but plausible-looking business answers create a persistent trust problem.
- Upside focuses on reconciling go-to-market data, definitions, and permissions so agents can answer questions about the business consistently.
- The speaker's central management principle is commander's intent: explain why an outcome matters so agents can make better implementation choices.
- Shared documentation, knowledgeable guidance, and second opinions are introduced as ways to prepare agents instead of relying on raw model intelligence.
agent-trustgo-to-marketcontext-engineering
Original description
Ask an agent how much pipeline you created in Q1 and it will happily take January to March and hand back a number that is wrong in exactly the way a right one looks. Alex Bauer's fix at Upside is a librarian. Before any agent acts it consults the documentation, the library of knowledge items about the business, and the schema of prior failed queries, then comes back knowing the fiscal year runs February through April and that pipeline only counts from stage two. Bauer opens with a bedtime story about a go to market team, a tidy pile of data, and a dragon, then admits every part of it is a real problem. The hallucination problem grew up into a trust problem, and his answer is to manage agents the way you manage people: give them commander's intent, the why behind the task, and stop them micromanaging themselves. Three examples carry it. Rebuilding Upside's website, two months of work for a product marketing team at his last company, failed in one shot until he wrote down the anchor assets first, including a product capabilities reference the agent compiled with citations back to every source system. The librarian is the second. The third is jury and judge for multi touch attribution: independent analysts each build an evidence cited opinion, a consensus judge weighs the reasoning rather than the conclusions, and a split verdict expands the jury. The bonus is agent tiers. Slack's new MCP client proved too weak to consult the librarian, and any model crowbarred into a per seat subscription probably is; important work needs a strong model with subagents, plan mode, and full MCP support. Speaker info: / alexdbauer https://x.com/alexdbauer https://upside.tech/ Timestamps: 0:00 - Otto, and a bedtime story to practice on 1:04 - Once upon a time: a GTM team, a data pile, and a dragon 4:06 - Every part of that story is a real problem 5:13 - The age of agentic GTM: technical enough to be dangerous 7:34 - Hallucination grew up into a trust problem 8:37 - Manage agents like people, and use commander's intent 9:27 - Example one: rebuilding the website from anchor assets 12:11 - Example two: the librarian 13:19 - Example three: jury and judge for attribution 15:15 - Bonus: agent tiers, and you cannot fix stupid