For this episode, I spoke with Sam Keen, founder of Altered Craft, author of Clean Architecture with Python, and a developer educator with 26 years of experience building production software, including leading generative AI work at AWS and working with engineering teams at Nike and Lululemon.
That history matters because Sam is not approaching AI-assisted development as a reason to abandon everything software engineers already know. He is asking which old assumptions still hold and which ones need reconsideration.
“You have to reset all your priors that you had as a software engineer.”
The best model is not the right model
In the early days of generative AI, using the most capable model available made sense. Every coding task sat somewhere near the edge of what the technology could manage.
That is beginning to change.
Frontier models can now complete many well-scoped engineering tasks reliably enough that using them for every stage may be unnecessary. Sam has been experimenting with what he calls personal tokenomics: deciding which parts of the work truly require frontier-level intelligence and which can be passed to a faster or cheaper model.
In his coding harness, Tilth, a frontier model, plans the work and another evaluates the finished result. The implementation between those two points can then be attempted by a less expensive model.
This is not simply a cost-saving exercise. It is an architectural decision about where intelligence is most valuable.
Reliable agents don't happen by accident
On November 21–22, we're hosting a 10-hour, hands-on Claude Certified Architect bootcamp with Puria Izady, covering agent architecture, MCP, Claude Code, multi-agent orchestration, and reliability.
You'll work through 6 practical labs, scenario-based exercises, and a 60-question practice exam to prepare for Anthropic's CCAR-F certification.
When 800 lines replace 130,000
That same instinct appears in Sam’s experiments with software dependencies.
He rebuilt a Flask application while questioning which third-party packages the project needed. The original dependency tree brought the application to ~210,000 lines of code. By replacing selected dependencies with around 800 lines of purpose-built code, the total fell to ~80,000.
He did not remove everything. Security-critical components remained with the specialists who had already built and tested them.
The exercise was less about declaring dependencies bad than asking a question developers often stop asking: where should the trust live?
Faster agents need better brakes
Sam’s experiments also exposed another problem. His implementation and evaluation models could enter a loop: one proposed a solution, the other rejected it, and the exchange continued while consuming more tokens.
The answer was not a smarter prompt. It was observability and a circuit breaker.
Tilth records every action taken by the model and its harness. If the evaluator rejects the work repeatedly, the system stops instead of allowing the agents to argue indefinitely.
As agents take on longer tasks at greater speed, that kind of instrumentation becomes essential. We cannot watch every decision as it happens, but we should be able to reconstruct what happened, see where progress stalled, and decide when the system must stop.
This conversation moves from model routing and dependency debt to context engineering, code review, and the skills newer engineers will need. But each subject returns to the same idea: better AI engineering does not come from using the most powerful option everywhere.
It comes from knowing where capability belongs, where complexity is justified, and where the system needs a boundary.
Watch the full episode for Sam’s take on choosing models, rebuilding software around AI-assisted development, and preserving engineering judgment while the tools change beneath us.
That’s it for this one. We’ll pick up the conversation next week.
Until then, keep building.
