Manage agents like new hires: three copyable trust blueprints -- commander's intent, the librarian, the jury.
Design Patterns for AI Trust: Juries, Libraries, and Agent Tiers — Alex Bauer, Upside.tech · Alex Bauer
17 min total·Actually worth watching closely: ~3 min·3 must-watch clips
- 0:12 – 3:46Listen
Opening fable: a quest for one number
He opens with a bedtime story: a Go to Market team sets off on a simple quest, wanting to know one thing -- how the business is doing -- and between them and the answer sit a simple pile of data, a dragon, and a series of guardians. Almost entirely narration, with nothing on screen you need to watch.
The real bottleneck for GTM teams is not that they cannot use AI -- it is that they sit too far from the data and too far from the answer.
A spoken narrative throughout, with the slide parked on the title page; fine to listen to while doing something else.▶ Jump to 0:12Speaker · Alex Bauer - 3:46 – 6:40Listen
Why GTM teams have nobody who builds
Compared with product and engineering, GTM teams have an extremely low density of people who build their own tools; the default toolbox was spreadsheets and slides. What AI does is make everyone technical enough to be dangerous -- an infinite supply of valedictorian interns handed to non-technical roles.
Putting the ability to implement in the hands of the person closest to the problem beats another layer of passing requirements down.
Carried entirely by what he says; no slide supports it. Catch the intern metaphor and you have the whole point.▶ Jump to 3:46Speaker · Alex Bauer - 6:40 – 8:40Listen
Hallucination has become a trust problem
Ask AI for a revenue report and it will not say it is unsure -- it gives you a wrong answer that looks entirely right. The nature of the problem shifts from the model makes things up to do you dare ship what it produced.
The dangerous answer is not the obviously wrong one; it is the one that looks exactly like the right one.
This stretch is a conceptual pivot -- he is talking, not demoing. The line about it never saying it is unsure is where the weight sits.▶ Jump to 6:40Speaker · Alex Bauer - 8:40 – 10:54Listen
Manage agents like people, and lead with the why
The talk's through-line: how you establish trust is the same for people and for agents. He borrows commander's intent from armed forces doctrine -- tell the agent why you want it to do this rather than only what to do, and the output quality changes noticeably.
If you take away one thing: always hand the agent the why. It works on humans too.
Pure argument with no visuals to decode, but worth hearing end to end -- it is the shared foundation for the three patterns that follow.▶ Jump to 8:40Speaker · Alex Bauer - 10:55 – 12:20Skim
Structure first: do not set it loose from the start
Define the rules before you free the execution: write down how your business world actually works, and package the permissions, metric definitions, known guides and the habit of asking for a second opinion into a single instrument you hand over. At the same time, require it to output every citation from the systems it touched plus the log of how it got there.
What you want is not a genius agent but a well-prepared one.
Slides start here: a bullet list plus a sample output carrying its citation trail. Skim for the structure and let the narration fill in the detail.▶ Jump to 10:55Speaker · Alex Bauer - 12:20 – 14:20Watch
The librarian pattern: check the archive, then answer
Before answering any business question -- say, how much pipeline did we create in Q1 -- the agent must first ask a dedicated librarian, which holds the company documentation, the knowledge library, and the schema of questions that have been answered wrong before. Which month the fiscal year starts, what a given metric really means: the librarian covers it.
Institutional memory should be a mandatory lookup step, not something you hope the model guesses right on the spot.
Nearly all the value is in that flow diagram -- who asks whom, what gets looked up, what comes back. Read the diagram and you can reproduce the pattern; listening alone tends to drop the middle steps.▶ Jump to 12:20Speaker · Alex Bauer - 14:20 – 16:00Skim
Jury and judge: what to do when there is no right answer
For questions like deal attribution that have no single correct answer, send out several mutually independent analyst agents to research separately and each return a cited-evidence opinion, then have a judge weigh them by reasoning quality; if there is not enough consensus, escalate and widen the jury.
Several independent researchers plus one adjudicator beats a single agent grinding on it forever.
The slide is a division-of-labor diagram, readable at a glance; the real substance is in the spoken lines about weighing reasoning quality and the escalation condition.▶ Jump to 14:20Speaker · Alex Bauer - 16:00 – 16:51Skim
Closing: you cannot fix stupid
Do not put important work on a bad harness or a weak model -- an AI product that makes its money on a cheap flat subscription has no margin left to run a smart reasoning model. He closes with the minimum bar for picking tools.
The floor of your tooling sets the ceiling of your results; no pattern rescues a weak foundation.
The final page is a bulleted floor checklist -- worth pausing to screenshot and check your own tools against, rather than listening to word for word.▶ Jump to 16:00Speaker · Alex Bauer