Random thoughts and working notes on building AI agents for production. Only this, and nothing more.
Local models are not free and subscriptions are not a rip-off. A plain look at what each person gains, what they give up, and the assumptions that decide when running AI at home pays off.
ReadAn agent makes a bad call in a regulated industry. Who's at fault: the company, the developer, QA, or the model provider? Field notes on where liability actually lands, the new AI insurance market, and why 'it passed testing' won't carry you.
ReadAn 800 line generated diff, a four minute skim, an approval that means 'this looks like code'. What review can still verify now that reading every line stopped being possible, and the tenth of the diff that still deserves slow human attention.
ReadAnthropic now marks every Claude output worldwide, including source code the EU explicitly exempted and prose you wrote yourself. The mismatch is structural: the law regulates content, the watermark lives in the sampler.
ReadAgent frameworks keep growing graph vocabulary. Model the run as a fold over an append-only event log instead, and durability, audit, human-in-the-loop, and distribution stop being separate features: they become consequences of one choice.
ReadEvery agent framework ships human-in-the-loop approval, and for demos they all work. The difference shows on resume: node-level replay re-runs the code before the gate and your audit trail drifts from what executed, while a step-exact log makes the trail identical to the execution.
ReadLetting AI write your code isn't a break from programming history, it's the next step in an 80-year retreat from the hardware. But this rung has one real difference: the abstraction underneath is probabilistic, not deterministic.
ReadA new study finds that the more you understand AI, the less excited you are to use it. The mechanism isn't fear, it's the death of magic.
ReadPrompt injection can't be escaped away like SQL injection. The fix isn't preventing it, it's surviving it: scope permissions in the infrastructure, gate the irreversible, and assume the attack succeeds.
ReadAn agent that works but costs four dollars a session doesn't ship. Three moves that cut agent cost without touching its logic: profile the tokens, cache the context, route the easy steps, and budget the rest.
ReadWhen an agent fails on step six, intelligence won't tell you why. A trace will. What an agent trace must capture, and the cold-reconstruction test for whether yours is real.
ReadAn eval suite that exists but isn't enforced ships regressions anyway. How to wire evals into your deploy pipeline as a gate, with baselines, noise-aware thresholds, and loud overrides.
ReadYour agent's confident diagnosis and its correct one look identical until you hold the ground truth. How to build the eval harness that tells them apart.
ReadThe agent loop everyone writes first dies with its process and double-charges on retry. Two structural moves fix it: split the decider from the doer, and make progress durable.
ReadThe agents that survive production aren't the most capable ones. They're the ones whose builders knew where competence ended and said so in code.
ReadA production AI agent is eight components, not a prompt. Here's the parts list, the failure each part prevents, and the four ways agents fail.
ReadMost AI agent pilots, 88% of them, never reach production. The gap isn't model intelligence. It's five engineering questions your demo can't answer.
Read