Notes
Short pieces about the methodology and architecture decisions behind the AI systems I ship — specs, evals, multi-agent orchestration, LLM integration, and the discipline of directing coding agents.
June 3, 2026
A cheap model can do 90% of the work
The default move is to point the biggest, smartest model at everything. It works in the demo and quietly bankrupts you at scale — because most of what an agent does isn't reasoning, it's mechanical, and you're paying genius wages to read a form. The fix is boring and worth ~90%: let a smart model plan, and cheap models do. Here's the economics, and the one architectural rule that makes it possible.
- agents
- architecture
- business
June 3, 2026
I stopped approving the agent's decisions. Now I watch them.
When I started building agents I approved every action — it felt responsible. It wasn't; by the thirtieth 'yes' I was rubber-stamping, which is worse than no checkpoint at all. Real oversight isn't touching every decision, it's setting the policy and watching the outcomes. Here's the shift from in-the-loop to on-the-loop, and the architecture that makes letting go actually safe.
- agents
- methodology
- ai-native
June 3, 2026
Grounding isn't a feature. It's a constraint.
An LLM is a guesser by design — it will always make things up, and you can't prompt that away. The only reliable fix is architectural: put a deterministic source in charge of the facts and demote the model to a rephraser that may never author one. 'Add RAG' isn't that. Here's the difference, and why it's the line between an AI that sounds confident and one you can trust.
- architecture
- agents
- methodology
June 3, 2026
One agent that does everything does nothing well
When an agent isn't good enough, the instinct is to give it more — another tool, more instructions, more context. That makes it worse, and it's measured, not a matter of taste. The fix is the oldest rule in engineering: Single Responsibility. One agent, one job, a few tools, a short context. A god-agent is a ten-thousand-line function in a trench coat — and it fails for the same reason.
- agents
- architecture
- methodology
June 3, 2026
Orchestration is the architecture now
Split the god-agent into ten focused ones and you trade a model problem for a systems problem: now they have to work together, and coordination is harder than any single agent. Most teams treat that wiring as plumbing. It isn't — it's the architecture, it's a distributed system, and it fails like one. Here's what orchestration actually is, how it breaks, and why you shouldn't reach for it until you can name the bottleneck.
- agents
- architecture
- methodology
June 3, 2026
Prompt engineering is dead. I never did it.
The industry spent two years hunting for magic words to whisper at the model. Now it's quietly declaring 'prompt engineering' dead and replacing it with context engineering and harness engineering. Here's the thing: those aren't new tricks — they're just engineering, the work real systems people were doing the whole time. Why the words were never the point, and what actually makes an agent work.
- agents
- methodology
- ai-native