Imagine a morning a few months from now. In your first meeting you look at a spec for the next feature: 40 pages of markdown, perfectly formatted. You do what senior people do – the equivalent of „looks good to me“ – and off it goes to the factory. Your second meeting is about last week’s incident: private information leaked onto the internet by a retry mechanism with an unusual fallback strategy. The customer is in the room. Someone asks you why the software made that decision.
You weren’t in the room when it was made. And the answer can’t be „the agents did it“.
That was the thought experiment Russell (Russ) Miles opened his keynote with at Agile Tour Vienna 2026. Russ has written The Sovereign Engineer and The Sovereign Apprentice, plus nine books for O’Reilly. His talk, he said, is „not a pro AI or negative AI. It’s pro human decision making.“
Hands-On AI Agent Engineering with Russell Miles, 16 December 2026 in Vienna
one day on your own codebase, so the quality of AI-assisted work matches the speed.
„The currency of software engineering isn’t code,“ Russ said. Not architecture diagrams either, not even value streams. „It’s decisions, decisions upon decisions upon decisions.“ And the one thing only humans can be held to is accountability for those decisions.
Two metaphors are back in the industry, and both work against that: the factory and waterfall. Russ once worked a twilight shift in a beauty products factory. Nobody there was looking for creativity or innovation. His problem with the factory metaphor is what it tells us to hide: the thinking and deciding that software engineering is actually about.
In 1996 Russ joined a defence contractor. In the interview they explained how they built software: nine months of specification and modelling in Rational Rose – and then they turned the handle and the code came out. He stayed eight years and, as far as he can tell, never delivered anything.
Now big specs are back. Spec-driven development promises great specs that an agent turns into software. To Russ, that „stinks of the waterfall„.
Jack Reeves described it in 1992: there is construction, which we can automate, and there is design – the thinking and deciding. The DevOps community has the building side covered. The interesting problem is design: exploring solutions, testing assumptions. That’s why Russ talks about software research and development. „The right answer is no software. Talk me up from there.„
Research needs small experiments with few variables, in the hands of users. A huge spec pushed through an agent pipeline is the opposite. „Huge experiments are useless. You are token maxing with 0 ROI.“
Waterfall’s real mistake, Russ argued, was batches: assumptions masquerading as decisions, implying yet more decisions. And we no longer turn specs into software with deterministic computation but with inference. Wherever the spec has gaps, the model makes decisions – and it tries to make you happy, not to annoy you.
Then comes the review. A big pull request on a system you haven’t been close to for a while, all tests green, and the reviewer does what most senior engineers do: looks good to me. Until someone asks why.
So which decisions do you need to stay close to? Russ uses a simple rule for what makes a decision big: „How difficult it is to fix it if you get it wrong.“ From that he builds four classes and asks his agents to find them in every spec:
In his example spec for a vet booking system, a log line is an undo decision. „Owners with unpaid bills cannot book“ is a candidate for the answer class – the kind of rule you will be asked about.
Russ thinks of the environment where humans and agents work together as a habitat. On his maturity model, level 1 is directing (telling the agent what to do), level 2 is commanding (reusable instructions). At level 3 you realise the environment matters as much as the prompt, and you build harnesses: the context agents need to do a good job. Level 4 brings teams of agents – „if you don’t get Level 3 right, level 4 is hell„. At the top he adds a new kind of agent whose only job is to help the humans.
Building that kind of harness for your own codebase is what the workshop on 16 December is about: a convention inventory, a version-controlled HARNESS.md with enforced constraints and two garbage collection rules in CI, all committed to your own repository. Join the workshop
Russ calls these agents Sentinels. They exist to help humans exercise judgement and take accountable positions:
Won’t that slow you down? Russ quoted a Ferrari engineer: the key invention that makes cars fast is brakes. „Your Sentinels are your brakes.“ Because software is not free – not in tokens, and not in the time and attention it takes to produce a quality product.
So does AI deliver waterfall? No, Russ said – but big batches do. What needs protecting is human judgement, the only accountable cognition in the room, and our ability to understand what we create. He pointed to Margaret-Anne Storey’s work on cognitive debt.
For every engineer, and especially for those starting out, he left the room with one question: „Can you defend, debug and evolve your decisions with and without AI?“
If you want to work on this with your own codebase: on 16 December 2026 Russ runs a one-day, hands-on workshop at TechTalk in Vienna for experienced software engineers and tech leads who already use AI coding tools. You work on why quality drifts in AI-assisted development, prompting as translation, a trust protocol for red-teaming AI output, and convention extraction. You leave with a convention inventory, a version-controlled HARNESS.md with enforced constraints, two garbage collection rules in CI – and a 30-day plan.
The slides from the talk are available on the Agile Tour Vienna website.