Blog
The engineering under an agent that does real client work. Mostly things that went wrong in production and what they turned out to mean — written while the evidence was still on screen.
Buying AI work · September 4, 2026 · 6 min
One number tells you whether AI is doing your client work or drafting near it: how much of each draft you rewrite before it goes out. If it is the same in month three as in month one, you bought a tool, not a colleague — and most vendors have never measured it.
Read →Cold email · August 30, 2026 · 9 min
Cold email rarely dies from one bad day. It dies from a fortnight of small mistakes — a ramp that looks like a leak, a sequence that retries dead addresses, one domain quietly carrying the whole day — and by the time you notice, the fix is a new domain and three months of waiting.
Running an agency · August 29, 2026 · 8 min
A junior on $55,000 costs about $100,000 in their first year once you count payroll tax, health insurance, recruiting, equipment and the fifteen days of your own time it takes to ramp them. Here is the full arithmetic, and the test for which work is worth hiring for and which is a process problem wearing a headcount costume.
Getting paid · August 28, 2026 · 7 min
Most overdue invoices are not a refusal to pay — they are an invoice that reached one person who is not the person who pays. A four-rung ladder where the escalation is in the specificity rather than the tone, the three rules underneath it, and the four things to fix before you fix the chasing.
Architecture · August 24, 2026 · 8 min
A bookkeeping service in our system is a JSON manifest, a folder of markdown, and an output schema. No module, no deploy, no engineer. The consequence that matters is not speed — it is that domain expertise stops being something only a programmer can add.
Architecture · August 23, 2026 · 8 min
Most agent frameworks treat a schema as validation applied after the fact. Treat it instead as the definition of done — the run ends the moment a valid result is written — and three unrelated problems disappear at once: unbounded loops, unresumable runs, and work nobody can grade.
Architecture · August 22, 2026 · 9 min
"Build a website for this business" and "decide the next step on this overdue invoice" are not the same job, and giving them the same tools, permissions and credentials is how an agent system becomes unsafe. Three shapes, and a build run that can never hold a send token.
Architecture · August 21, 2026 · 8 min
Real service work is not a single prompt. A weekly visibility report is twelve independent measurements and one aggregation; a month-end close is strictly sequential and stops dead without a bank statement. Both need a run that can suspend itself as a row and resume days later.
Architecture · August 20, 2026 · 7 min
Row locks stop one job running twice. Nothing in a typical system counts how many times one human hears from you. Three correct schedules, three correct runs, three emails to the same client before lunch — and no bug to find, because nothing is broken.
Product · August 19, 2026 · 8 min
A finished agent run has exactly three honest fates: deliver, ask, or hold. The fourth — ship whatever came out — is the one that puts a serialised object in front of a paying customer, and it is almost always a single fallback expression away.
Product · August 18, 2026 · 7 min
Bookkeeping needs the bank statement. A hire needs the role brief. Most agent demos quietly assume the input is already there, and every real engagement stalls on the moment it is not. Refusing well is a product feature, and it has a shape.
Engineering · August 17, 2026 · 7 min
If a run can suspend for three days waiting on a client, every piece of state that resumes it has to survive a deploy, a crash and a second replica. The rule is uncomfortable and simple: if losing it would park a run forever, it is not allowed to live in memory.
Agents · August 16, 2026 · 9 min
A scripted synthetic client can only prove your transitions exist. The loop that earns a retainer is different: v1 goes out, the client objects to one thing, v2 comes back, and they ask the only question that matters — did they fix the thing I said? That needs memory.
Research · August 15, 2026 · 8 min
The most useful distinction in AI visibility, and almost nobody makes it. Being named from training data and being retrieved as a source for one answer are different phenomena with different time constants — and only one of them is work an agency can sell.
Research · August 14, 2026 · 12 min
Protocol 4.0: two planes, never averaged. Judgment head-to-head — Mycel 21/21 (100%) vs a naive unscaffolded lower bound 11/21 (52%). Architecture behavior-graded 42/42 on real kernel modules; peer coding harnesses are capability-absent, not scored 0%.