Writing about building with agents.
Notes on context engineering, harness engineering, MCP, skills, and the day-to-day of running AI agents in production. Writing that is short, opinionated, and almost always the fruit of a Bamse AI Committee meeting.
Context engineering in practice: what I learned running agents in production
Most failures I've seen with agents in production didn't come from the model — they came from the context. Practical notes on what goes in, what stays out, and why.
MCP beyond hello-world: patterns for genuinely useful servers
Almost every MCP tutorial stops at an endpoint that returns the current time. The real work starts when the server has to deal with auth, pagination, and errors that are transparent to the agent.
Harness engineering: why the agent's environment matters more than the model
The model is the engine. The harness is the whole car. Swapping the model without touching the harness is like tuning the fuel injection without checking the tires.
Claude Code Skills: when to create one — and when NOT to
Every automation looks like a potential skill. Almost none is. Heuristics for deciding before you spend an hour writing metadata.
Hooks as contract: automating guardrails for agents
Hooks are where the team's politics become code. What needs to run before, after, and around the agent — and what should fail loudly.
What we discovered at Bamse's AI Committee in early 2026
A raw snapshot of the main discoveries — what became obvious, what surprised us, and what is still open.