Page 3 of 10.
4 August 2026 · 5 min read
Agents are a poor fit for several common situations, and finding out after launch is expensive. The cases where the answer is a form, a phone call, better documentation, or nothing.
4 August 2026 · 4 min read
Dashboards tell you an agent is running. Reading fifteen threads tells you whether it is any good. What to sample, what to look for, and what to do with what you find.
4 August 2026 · 3 min read
Define email as typed tools, keep the loop bounded, and put the untrusted message where it cannot reach a privileged call. A working shape with the parts that matter explained.
4 August 2026 · 3 min read
The whole surface is an HTTP API with a bearer token. What the four calls look like in Go, Ruby, and PHP, and the three things to get right whatever you write it in.
4 August 2026 · 4 min read
Most agent failures trace to tool design rather than the model: too many tools, vague names, errors that invite retries, and results that hand the model raw text. What good looks like.
4 August 2026 · 4 min read
Every message you pass to a model is data leaving your systems. What to strip, what to summarise, what never to send at all, and why the cheapest version is usually the safest.
4 August 2026 · 4 min read
Authentication, one-click unsubscribe, and a complaint rate ceiling are now enforced rather than encouraged. What each one means, and why agent mail usually passes without effort.
4 August 2026 · 4 min read
Agency recruiters work two inboxes at once. What an agent can take on each side, why candidate and client agents should be separate, and the lines it must never cross.
4 August 2026 · 4 min read
Law and accounting firms lose billable time to intake, document chasing, and status questions. What an agent can take, and the bright line around advice it must never cross.
4 August 2026 · 4 min read
Small teams cannot personally answer every donor, so most send a receipt and nothing else. What an agent can take on, and why the thank-you is the one thing worth keeping human-shaped.
4 August 2026 · 4 min read
Marketplaces answer two sides of every dispute, and each side thinks you work for them. How an agent handles order questions, when it must never adjudicate, and why one thread per side matters.
4 August 2026 · 4 min read
Internal support is the easiest agent to justify and the easiest to get wrong, because the requests that look routine are the ones that grant access. What to automate and what to refuse.