Essays on AI agents, the systems behind them, and building companies

What broke the prompt cache
Filtering the tool list on every turn broke the prompt cache: 74,000 tokens written per turn, none read, and 44 seconds to the first token. The fix was to stop changing the list.

Agents are the primary users
Tools built with an agent as the first caller: reachable from a shell, from another agent and from code, with compact machine-readable output, failing closed.

Hard limits are governance
A guideline needs someone awake to enforce it. A number in the rules file stops an agent at 3am, and it is changed there, in writing, when the evidence says so.

Token economics
Token cost has to be measured like latency: per account, per agent and per machine, and checked against the invoice. A monthly total arrives too late to act on.

Model routing in practice
Model routing is decided by the gaps a vendor leaves: one gateway under everything, a circuit breaker for the bad day, and each model on the job it does well for its cost.

Debloat is a feature
Debloating moves complexity to where a person never has to look. With agents building all day, it holds only when the ceiling is stated as a number or a rule.

Durable state for agents
Agents crash and machines reboot. State in a local SQLite file, every run recorded and idempotency keys let a process end without losing or doubling work.
2025
3 essays