Agent BoundariesDraw the Line, Then Let Go

The argument

Models keep getting more capable, but capable isn't the same as useful. This series tries to answer one question: how much autonomy does a given use case actually need? My approach is to draw the line first: which level is enough, where the boundary sits, how to stop it when it goes wrong. Once the line is clear, I let go one level at a time.

Parts, in reading order

1. Part 1FoundationsAI product · 23 Mar 2026 · 6 min Moving AI Agents from Hype to Engineering Reality: Finding Your Sweet Spot in the Architectural Spectrum Don't jump into multi-agent systems just because they look advanced. Let's talk about the five levels of Agents and why in 2026, defining 'boundaries' is more critical than enhancing 'capabilities'.
2. ↓First, get the two simplest levels right
3. Part 2PrecisionAI product · 22 Apr 2026 · 19 min L1 & L2: Teaching AI to Be Precise, Not Just Prolific When a system only needs to answer, a precise answer is worth more than a long one: most of the work in L1 and L2 goes into retrieving the right passage, routing each request the same way every time, and refusing when nothing supports an answer. Where that stops working is how you tell whether you need an Agent at all.
4. ↓Let the model choose the next step
5. Part 3ActionAI product · 24 Apr 2026 · 12 min L3: Let the Model Choose the Next Step, but Let People See Every Step From L3 on, the model decides what to do next. This post covers how it asks follow-up questions and calls tools, and why every step should be visible to the user, with a person confirming before anything is written or sent.
6. ↓When goals conflict, who decides
7. Part 4CollaborationAI product · 19 Sep 2026 · 5 min L4: Before You Add Another Agent, Decide Who Has the Final Say You need L4 only when a task holds several conflicting goals and someone has to make the final call. Before adding agents, decide which goal each one speaks for, who reviews, and when to stop.
8. ↓How to tell the system actually got better
9. Part 5EvaluationAI product · 28 Oct 2026 · 5 min Evaluation: However Fast Agents Change, First Ask What Problem You're Solving Agent products turn over every few months, and they're given more and more access. This last post in the series is a short note rather than a long argument: however the technology changes, first get clear on the business problem, which level of the technology you need, and how to make the results reliable.

Still open

1. 01Once an agent leaves a fixed workflow, how do you draw the boundary of what it may decide on its own?
2. 02When several agents each speak for a different goal and pull against each other, where do things most often go wrong?
3. 03After launch, how do you use what the system actually does to improve it, without making every step wait for a human review?

1. 01 Note Why Does Your AI Always Agree With You? On AI SycophancyAI sycophancy, the tendency of mainstream large models to please the user with flattering, agreeable answers over the facts, starts in reinforcement learning training. Five habits in how you ask can get you more objective, sharper opinions out of AI. 5 Jun 2026 · 7 minRead →
2. 02 View Vibe Coding Gets You There Faster — It Doesn't Choose Where You're GoingWhat Vibe Coding actually reduces is the cost of making ideas visible — turning vague descriptions into something that can be seen and touched. What it doesn't reduce is the cost of judging whether something is worth building in the first place. The faster the tools, the more easily a wrong direction gets amplified. And the people who can translate between business logic and technical implementation? They're becoming rarer, not more replaceable. 25 Apr 2026 · 6 minRead →
3. 03 View Is Your Bulleted List "Too AI," or Just Thoughtfully Structured?Redefining the role of structure in professional communication in the age of Generative AI. 18 Mar 2026 · 2 minRead →