# Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers Alex Bauer, Upside.tech | AI Engineer World's Fair 2026 | 17:09 Source: https://www.youtube.com/watch?v=YZQsWVeN3rE Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/design-patterns-for-ai-trust-juries-libraries-and-agent-tiers Published: 2026-07-11 Tags: agents, context-engineering, multi-agent, rag ## TL;DR - AI agents need the reason behind a task, along with documentation and business context, because a fluent wrong answer can look correct. - Anchor assets let an agent build from a defined structure, while citations let people check the business claims it uses. - Independent analysts and a consensus judge can handle attribution questions where no single empirically correct answer exists. ## Summary Alex Bauer argues that the main AI problem for go-to-market teams has become trust. Agents make it easy for people without engineering backgrounds to build software, but they can return answers that sound right while using the wrong definitions or data. Bauer's practical answer is to manage agents like people: explain the purpose of the task, document how the business works, and give the agent ways to ask for context and second opinions. He shows this through three examples. A website rebuild worked after he created anchor assets and a cited product capabilities reference. A librarian gives agents just-in-time business definitions before they query data. A jury and judge workflow uses independent analysts and a consensus judge for multi-touch attribution. He also argues that important work needs capable models and harnesses with features such as subagents, plan mode, and full MCP support. ## Key ideas ### AI has made more go-to-market workers capable of building software [05:27](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=327s) Bauer says go-to-market teams historically relied on spreadsheets and PowerPoint rather than code. AI changes that by giving people an "infinite supply of valedictorian interns with computer science degrees" to implement ideas they often understand better than an engineer from another team. This speeds up iteration, but it also lets nontechnical users send agents directly at business problems. Bauer describes himself as "technical enough to be dangerous," and says AI now makes all of us fit that description. The resulting systems can be useful, but they need guidance before people trust their output. ### A fluent wrong answer has become a trust problem [07:23](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=443s) Bauer says the familiar hallucination problem has grown into a trust problem. Ask Claude to report revenue and it will not necessarily say that it is unsure. It may return a wrong number that looks exactly like a correct one. He is skeptical that most teams can spend their time chasing perfect prompts or the newest technical optimization. His practical approach comes from working with people: establish trust by giving agents context and clear direction. He frames the rest of the talk around applying human management habits to AI systems. ### Commander's intent gives an agent the purpose behind its task [08:40](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=520s) Bauer's main prompting advice is to use commander's intent, a concept from armed forces doctrine. Instead of only specifying an action, tell the agent why you want it done. He says this improves results for both agents and people because humans generally dislike micromanagement. Agents can also micromanage themselves because they were trained on human material. Telling Claude to improve itself may produce more self-directed process than the task needs. Bauer says to pull it back to the "why" when that happens. ### Anchor assets give an agent a business structure before it writes [09:27](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=567s) Bauer tried asking Claude to rebuild Upside's entire website from a collection of source materials. Even with plan mode, the one-shot approach failed. He then created a maintained list of company anchor assets and defined the business structure first. One asset is a product capabilities reference. AI compiled it, but each capability includes what it does, why it matters to different personas, and citations back to the connected source systems. Bauer can follow those citations to check the claims. The resulting website content maps back to the reference. His instruction is simple: define the structure first, then turn Claude loose. ### A librarian gives agents just-in-time definitions before they query data [12:21](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=741s) Bauer's librarian sits between a user's question and the agent's data work. If someone asks how much pipeline the company created in Q1, an agent might assume that Q1 means January through March and count records by created date. The librarian first consults company documentation, knowledge items, and the schema of prior failed queries. In Bauer's example, it tells the agent that the fiscal year runs from February through April and that pipeline means opportunities at stage two or later. The final answer includes citations, so the result is grounded in definitions the business has already established. ### Independent analysts and a judge handle attribution ambiguity [13:19](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=799s) Bauer applies a jury and judge workflow to multi-touch attribution, where there is no empirically correct answer for every go-to-market question. For an Acme deal, independent analyst agents examine the data separately and produce evidence-cited opinions about attribution credit. A consensus judge treats those opinions as inputs rather than facts. It weighs the quality of each analyst's reasoning and produces a final result. If the analysts do not reach enough consensus, the judge expands the jury. Bauer says this is better than one person or agent repeatedly perseverating on the same research. ### Important work needs a capable model and harness [15:15](https://www.youtube.com/watch?v=YZQsWVeN3rE&t=915s) Bauer's bonus point is that teams cannot repair a weak setup with wishful prompting. He tried Slackbot's new MCP client with the Upside librarian and found it too unintelligent for the job. He generalizes that AI features added to per-seat subscription products may not leave enough margin for a strong reasoning model. For important work, he recommends at least a more capable agent tier with a powerful model, subagents, plan mode, full MCP support, and file editing. A basic ChatGPT web interface should not be expected to produce the same result. ## Notable quotes - "The thing about AI is it makes all of us technical enough to be dangerous." (06:51) - "It says, "Here you go." And it gives you a wrong answer that looks exactly like being right." (07:43) - "Use commander's intent when you prompt." (08:58) - "Define the structure first, and then turn Claude loose." (11:25) - "My job is not to do research on my own." (14:47) ## Tools & references mentioned - Upside.tech - Claude - Slackbot - MCP - ChatGPT - Opus - Consult Persona Bench - Jeff from xAI ## Who should watch - You are putting AI agents in front of revenue, product, or marketing data and need answers that use your company's definitions. - Your team can build with AI but has started getting polished outputs that are difficult to verify. - You are designing multi-agent research workflows for questions where evidence can support several reasonable conclusions.