Your agents can work all day. Can you keep up with reviewing it?
Chris Lucian is director of global software development at Hunter Industries and one of the people who started mob programming, where three or four developers work on the same code at one computer at the same time. At Hunter, he says, a mob of four finished more than four times as many Kanban cards as the same four people did working separately, with better quality and more frequent releases. The practice is now called software teaming, after Amy Edmondson’s term “teaming.” Chris has a degree in machine learning, hosts a podcast about mob programming and speaks at conferences around the world. His talk on agentic AI, “kite flying and plate spinning,” is going to Australia and Japan next. Hunter now has eight mobs working with agents.
His main argument: as agents do more of the work, your job turns into handing off everything that doesn’t need a person’s attention, and most people haven’t had much practice. Working with agents means delegating to machines you don’t fully understand. Managers who already delegate start ahead. Lifelong individual contributors, the “rock star” producers who never wanted to manage, struggle most. They want to see exactly what they would have typed, so they keep their prompts small. Chris tells managers to send those people to delegation training. For business users he describes a ladder. First, ask mode: use the AI as a thought partner (“pressure test my thinking” before a meeting), not as a source of facts. Next, connect read-only tools, like an Outlook connector that drafts a reply to every email for you to send. Then agent mode, such as having it write spreadsheet formulas instead of numbers, because an LLM predicts digits rather than calculating them.
When agents are doing the work, people reviewing it become the bottleneck. “Everything is determined off of our ability to give critical feedback right now to the models,” Chris says, and that’s as true for executives as for developers. His own agent loops run 24 hours a day and send video updates to his phone. He critiques them and sends them back, often without opening his laptop. He works on the bottleneck from two sides. The first is giving people less to read. His agents answer in dense bullet points, and a Risk-Aware Commit Notation MCP he built with Jay Bazuzi splits agent work into small commits that are easy to read, each labeled with how risky the change is, so reviewers know where to look hardest. The second is reading faster. For over a decade he’s used an open-source speed reader that flashes words at 500 a minute. And whenever a review catches a mistake, the fix goes back into the agent’s skills or instructions so it doesn’t happen again.
The rest of the discussion
- The TDD debate: do agents really need to write the test first, or are they just “gaslighting” you? In test-driven development you write a test for one business rule (say, only finance can start a transaction), then write only enough code to make it pass. People argue online about whether agents need that order, or whether they say they’re doing TDD while writing every test and all the code in one go, which Chris calls “AI gaslighting.” In his experiment, a small local model, Gemma 4 (it fits in 8 GB of RAM on a Costco computer under $2,000), is forced through a strict step-by-step process. Left alone, it can’t do a simple tic-tac-toe exercise test-first. Inside his system it does quite well, gaslights less and uses fewer tokens. He likes TDD but hasn’t decided whether agents need it. Unit tests, he says, are not optional. When Sachin mentioned developers who say TDD would raise their token spend 4 to 5x, Chris’s reply was: until you hit a million tokens and it starts making mistakes everywhere.
- Why vibe-coded apps fall apart: they outgrow what the model can hold in its head. Frontier models now have 1-million-token context windows, enough for the whole Lord of the Rings trilogy plus The Hobbit. A CEO vibe-codes something that works well and hands it to engineering, and it gets worse once the code base grows past what the model can keep in working memory. After that, nobody knows how it works. His quick test for vibe-coded versus properly engineered: count the unit tests.
- Mutation testing gives agents “wind to the kite” without burning tokens. A tool makes small automatic changes to the code. If no test fails, that logic isn’t covered. It can run for hours without spending a single token and produces a list of gaps for an agent to fix. The agent starts it, waits, and then picks the work back up. Chris says this is how his round-the-clock loops turn out well-factored code on a consumer plan without spending much on tokens.
- Local vs. frontier models: why smaller models get confused more often. Chris moves between local models (Gemma 4, Qwen, which he calls “really, really great already”) and frontier models to see where they differ. Frontier models now label commit risk correctly pretty much every time. His explanation for the difference is that larger models keep concepts further apart internally and smaller ones pack them closer together, so they mix things up, like calling a risky feature a harmless refactor.
- Are LLMs hitting a plateau? Chris thinks so; Doyle isn’t convinced. Chris points to an MIT paper, which he heard about from a YouTube review. It argues that frontier models rarely confuse tokens anymore and that every doubling of their representational room takes about six times the hardware, which now means billions of dollars per doubling. He compares it to the AI winter, when funding dried up after a paper suggested neural networks couldn’t solve a basic logic problem. He expects the next big step to come from something other than transformers, probably from universities, and much sooner than the AI winter’s 30 years. He admits he’s reading between the lines (“take from that what you will”). Doyle brought up the bitter lesson and argued that even if models stall, the application layer will move so fast that business users never get a break. Sachin asked whether the labs are spacing out releases because they can see the ceiling. Chris agreed they may no longer be able to predict when the next model will arrive. All three agreed the application layer is having a renaissance right now, and Chris sees one good thing about a slowdown: a breather for people still learning to work with agents.
- Ask, plan, act: Chris’s robot analogy, and why adoption moves at different speeds. Ask mode is C-3PO following you around and answering questions. Plan mode is a restaurant robot that takes names and manages the wait list. Agent mode is the cat-faced robot at a pho restaurant that actually brings your food. He explains adoption with diffusion of innovation theory: agentic AI has already crossed the chasm, and the late majority is mostly people who resist change or have always worked on their own. Going from “ask permission” to “just go” happens gradually, as people get comfortable.
- Chris’s reading list. The AI-Driven Leader, from his executive book club, for ask-mode techniques like interviewing yourself and staying in the driver’s seat instead of handing off your thinking. On AI bias, The Alignment Problem and Weapons of Math Destruction. He says Google once held back word-embedding technology because it produced associations like “doctor plus woman equals nurse,” while OpenAI, with nothing to lose, shipped anyway.