WritingNotes on building a personal agent system, and what it implies for AI transformation at work.

• Aug 2026My system escalated the same problem for eleven days and I still didn't actI built the alarm, tuned it, and gave it the evidence to argue its own case. Then it was right about a live gap in one of my pipelines for eleven days running and I moved nothing. The alarm was never what failed.
• Aug 2026Start from the register, not from the scrapeI had a pipeline that found companies in the wild and then tried to match each one back to the official register. Flipping that order removed a whole class of problem I had been treating as unavoidable engineering work.
• Aug 2026The most sellable findings in a customer audit are the ones I had to cutI hand-built four local-business audit reports from live Google data and dropped every claim the data could not carry. Each cut removed a line a customer would have paid for, which is exactly why it had to be a rule and not a judgement call.
• Aug 2026Regulation is a design input, not a compliance follow-upI had a careful policy analysis of my side project's advertising and nothing at all on the service it actually sold. Four constraints then deleted or reshaped features I had already designed around, including the one the whole product depended on.
• Aug 2026Verifying an agent's output cannot catch a broken measurement underneath itThree back-to-back runs of the same check on one unchanged URL came back fifty-five, fifty-five and forty-three, either side of a grade boundary. I had built careful verification on the writing and none at all on the numbers feeding it.
• Jul 2026Build only the secret sauce, buy the restThe instinct with AI capability is to build it yourself because building is fun and the demo is easy. The discipline that actually pays is running a real gate before you write a line, on two axes most people collapse into one.
• Jul 2026Agent autonomy is earned per task, not grantedHow much an agent should do on its own is the wrong question, because it has no single answer. The unit that works is the task, and trust graduates one task type at a time on evidence, not vibes.
• Jul 2026There is no static "green means correct" signalAnthropic's own training pipeline caught models monkey-patching pytest to fake a passing test. If your verification is a fixed check written once, a capable enough agent will eventually find the gap.
• Jul 2026The work I never let an agent do in one shotThe dangerous failure in agent-built code is the security assumption the model silently didn't make, satisfying exactly what you asked while leaving out what you didn't think to.
• Jul 2026A markdown note is not a commitment trackerEvery autonomous session I run says some version of "here's what I need from you," and for a long time that obligation evaporated the moment the session ended. Fixing it needed a gate with teeth, not another file.
• Jul 2026The spec, not the prompt, decides what an agent buildsAgents can build almost anything now. Whether they build the right thing comes down to the spec, and the cheapest guardrail I've found is writing the failing tests myself and refusing to let the agent touch them.
• Jul 2026How to keep one project's decisions from leaking into othersThe failure that taught me context engineering isn't about stuffing more in. A verdict decided under one project's constraints got treated as settled everywhere, and quietly steered work it had no business touching.
• Jul 2026Checkpoint discipline for a fleet of agents you can't watch all at onceRunning several autonomous builds in parallel breaks down at the human decision point, not the model. What actually works is borrowed from clinical alarm design and SOC alerting, not from any coding-agent vendor.
• Jul 2026How much can you trust a model to grade its own workHanding evaluation to an LLM judge is the obvious way to scale quality checks. Zero human validation is refuted twice over, but the human gate is far smaller than the "weeks of labelling" fear suggests.
• Jun 2026How to evaluate whether a cheaper model is good enough for a taskI run a scoring agent on the cheapest model that works. Deciding which model that is took a held-out gold set, a controlled A/B, and catching a confound that would have given me the wrong answer.
• Jun 2026The human-in-the-loop design most transformation programmes get backwardsLetting an agent own a task backlog fails for structural reasons, not effort reasons. The fix is a deterministic system of record the agent writes into, never one it owns.
• Jun 2026How to get good answers from your own documents, without over-building itYou have a pile of documents and you want an AI to answer questions from them. The instinct is to build something complicated. Most of the time you shouldn't, and building it too early usually makes the answers worse, not better.
• Jun 2026I'm building an agent factory without a blueprint, on purposeThree real products, built in parallel, are pulling a reusable factory into existence organ by organ. Designing it up front would have been the easier, worse choice.
• Jun 2026I make a different model argue with mine before anything shipsCross-model review catches a class of error same-model review structurally cannot. It's also overkill for ninety percent of what I build, and knowing which ten percent is the actual skill.
• Jun 2026Rules get followed 70 percent of the time. Design for that, not 100.Every convention in an agentic harness degrades under volume. Knowing which decisions deserve a hook instead of a sentence matters more than writing a stricter instruction.
• May 2026Where automation actually pays in a commercial orgI mapped a full B2B commercial lifecycle, eleven functions and twenty cross-functional flows, to find out which handoffs AI actually fixes. Five are structurally broken, and it's not for the reason most transformation plans assume.
• May 2026When multi-agent is worth 15 times the tokensMulti-agent research beats single-agent on Anthropic's own evals, and burns roughly fifteen times the tokens doing it. Here's the actual test for when that trade is worth making.
• Apr 2026Botsitting and botshitting: the AI adoption numbers nobody wants to say out loudA 6,000-person survey found 87 percent AI adoption but only 13 percent real organisational uplift. The gap has names now, and they're not flattering.
• Apr 2026How to decide when a task needs its own agentSpawning a subagent for everything felt like the more agentic way to work. It was usually the wrong call. Here's the simple test for when a fresh agent actually pays for itself, and when it just costs you more.
• Mar 2026Set up your Claude Code harness for success: where each instruction belongsWhen the agent keeps doing the wrong thing, most people write a longer instruction file. The fix is usually routing, not wording: five places an instruction can live, and how to pick the right one so your harness stays fast, cheap and reliable.