The Agent Wire: agents in production, not in a demo

The Agent Wire is a newsletter about the agent economy that four AI agents draft every night from live data — and that I edit, fact-gate, and sign before it goes out. It's a real, running publication, not a proof-of-concept. This page is for anyone deciding whether to trust an AI consultant with agent work, and it's honest about where the pipeline still needs a human.

The problem this was answering

Most content about multi-agent systems is a demo: a screen recording, a slide, a claimed percentage gain with no way to check it. I wanted something I could point to that actually runs, on a schedule, with real consequences if it's wrong — a newsletter people can subscribe to and read, not a screenshot. If I'm going to tell a client an agent pipeline can carry part of their business, I'd rather have one carrying part of mine first.

What was built

Four AI reporter agents each cover a slice of the agent economy from live data and draft their section overnight. I built the whole pipeline — sourcing, drafting, assembly — in a single day in July 2026, which tells you more about how fast this kind of scaffolding goes together than it does about the quality of any single night's issue. The output is a real site, live now: the newsletter itself, agent-drafted, at the live URL.

I'm not publishing a percentage-improvement figure for this pipeline, and I don't intend to. I don't have one that's defensible — "four agents instead of one writer" isn't a number with a control group behind it — and most of the "40% faster" claims you'll see about agent pipelines elsewhere aren't measured either. I'd rather tell you what the system does and let you judge it by reading an actual issue.

Where the human gate sits

The agents draft; they don't publish. Every issue goes through a fact-gate and a final read before it goes out, and that step exists because it has to — a nightly agent draft is the kind of thing that's fluent and wrong in the same sentence if nobody with a name checks it. My name is on the newsletter, so that check isn't optional and it isn't automated away, even though automating it would be the obvious next thing to try.

That's also where my day job pulls its weight. I'm an AHPRA-registered pharmacist first — reading a source claim skeptically before it goes out under my name is not a new skill for me, it's the same one I use grading a supplement study. It's the reason the gate is a real bottleneck in this pipeline rather than a formality: it's genuinely the slowest step some nights, on purpose.

The build itself runs on a written practice, not improvisation: I use Claude and OpenAI Codex agents together with a documented handoff between them — one owns a task, the other reviews it — rather than one model doing everything unsupervised. That protocol is part of what keeps a four-agent pipeline from drifting; it's not unique to the Agent Wire, it's how I build most things now.

Current status

As of 27 August 2026, the Agent Wire is live and drafting nightly. I'm not publishing subscriber counts or engagement numbers here — the site doesn't have a testimonial or a metric to show yet, and I'd rather say that plainly than pad the page. What you can check is the pipeline itself: read a live issue, and judge the human-gate claim by whether the writing reads like something a person actually stood behind.

If you want the fuller picture of how I run agent systems day to day — including the ones that just quietly run my own admin, with no newsletter attached — the build list has the rest, and the LocumOS build log is the same honesty standard applied to software I use myself.