BONUS AI Is Not Just Typing Faster With Marcus Hammarberg

BONUS: AI Is Not Just Typing Faster With Marcus Hammarberg

In this BONUS episode, Vasco Duarte and Marcus Hammarberg explore what changes when AI agents move from writing small functions to creating whole features, running through tickets, and even improving their own instructions. The conversation keeps coming back to a simple but difficult question: if agents make software delivery faster, what should humans use that speed for?

When the Agent Writes the Feature

"What I realized was that the agent could now write the code from a prompt that I gave it, and it was good enough that I didn't have to touch a lot of it."

Marcus starts with the moment that changed his own practice: in December 2025, he used an AI agent to build a feature for a flow metrics app based on Jira change logs. Before that, AI coding felt like asking for small functions and then cleaning up the result. This time, the agent could take a feature-level prompt and produce something useful. That shifted the question from "can AI help me type code?" to "what happens to planning, quality, and ownership when implementation becomes dramatically faster?"

Slowing Down Enough To Think

"You end up not only not having seen the code, but not maybe even know what you wrote."

Vasco connects Marcus' experience to an earlier episode where Marcus had to learn to program after breaking his hand. Back then, the lesson was that typing was not the bottleneck. Thinking was. Now AI has made "typing" even faster, and Marcus warns that this creates a new danger: teams can approve suggestions, specs, and generated code without really understanding the intent they gave the system. In that world, slowing down becomes part of the human job.

From Developer To Robot Instructor

"With the robot in place, what are you doing now? You're giving the robot instructions, and you're verifying the result. This is exactly what we are doing with the AI now."

After rereading The Goal, Marcus started to see AI agents as the software equivalent of factory robots. A welder used to make the weld directly. With robots, the human writes instructions and verifies the result. Software is entering the same kind of transition. The open question is uncomfortable: can someone reliably verify AI-generated software without being a strong developer first? Marcus does not pretend to know the answer, and that uncertainty is part of the value of the conversation.

Harnesses, Feedback Loops, And Agent Teams

"A really good harness should also have some deterministic checks in the feedback mechanism."

Marcus describes teams experimenting with "lights-out" agent factories: a project manager agent receives intent, delegates to sub-agents, the agents move work across a Kanban board, and they even run retrospectives to improve their own instructions. That sounds futuristic, but Marcus brings it back to Agile basics. If prompts and guardrails are non-deterministic, then the harness needs deterministic feedback: tests, linting rules, functional checks, non-functional checks, architectural rules, and explicit slices of work that humans can test before moving on. The lesson is not to abandon Agile. It is to make feedback loops more concrete because the work is moving faster.

The Human Side Of Losing The Easy Wins

"I don't feel the same kind of ownership when I have prompted my way to a good application as I did when I wrote it."

The episode also names a quieter transformation: developers may lose the small dopamine hits that came from solving tiny coding problems, seeing tests pass, and refactoring by hand. If agents handle the easy work, humans may be left with harder, slower problems: intent, architecture, product value, safety, and trade-offs. That changes the emotional shape of development. Marcus and Vasco explore what happens when a developer is no longer only a coder, but the manager of an agent team.

Stop Measuring AI By Output

"Awesome, are you selling percentages?"

Marcus brings the conversation back to Theory of Constraints: throughput is not how busy the robots are, how many pull requests were merged, or how many tokens were consumed. Throughput is economic value. Vasco and Marcus connect this to product thinking: if AI lets teams produce more features, that only matters when those features create better outcomes for customers and the business. Otherwise, AI becomes a faster feature factory. As Marcus puts it, if people do not understand the "why" and the metric they are optimizing, AI is truly just typing faster.

What To Read Next

"It starts and stops with The Goal."

Marcus recommends starting with The Goal by Eliyahu Goldratt, then applying its factory lessons to software, flow, and AI agents. Vasco also mentions Ari-Pekka Skarp's Scrum Master Toolbox episode, Philip Su's work on dark factories, and Spec Kit. Marcus points listeners to his writing at marcusoft.net, where he is exploring AI as software development's new set of robots.

About Marcus Hammarberg

Marcus Hammarberg is a Swedish agile consultant, Kanban practitioner, and author of Kanban in Action and Salvation: The Bungsu Story, about using Lean, Agile, and Kanban to help an Indonesian hospital recover from crisis. Previous Scrum Master Toolbox episodes: Marcus page. LinkedIn. Today he explores agentic software development in practice.

You can link with Marcus Hammarberg on LinkedIn, read his writing at marcusoft.net, and find his previous Scrum Master Toolbox episodes on the Marcus Hammarberg guest page.