AI agent infrastructure engineer · Alludium · formerly Principal Software Engineer, Dell · Cork, Ireland
Anyone can get an agent to write software now. The hard part is the system around the model: what it sees, what it is allowed to do, how its work is proven, how it is steered without hand-holding, and how what it learned survives to the next session. That is the work I do.
Six things I’ve learned doing itTeam practice, personal tools and one experiment, each labelled. Short, specific, and honest about what did not work.
In October 2025 I brought spec-driven development into the platform team I work on. A year later we retired it, and kept the parts that had earned their place. Both were the right call at the time, and the gap between them is the most useful thing I learned last year about working with AI agents.
Team delivery practice at AlludiumThe Golden Runner is not a clever piece of software. It's a spare checkout of the platform on my Mac that agents can borrow to run the real app and click through it before they open a PR. I'm writing about it because of the rules around it, and the rules exist because of a thing every agent does.
My own development fixture, used dailyIn May 2026 I started a Git repository called Work Memory. It's a folder of Markdown files that my agents read and write, and five months later it has about eleven hundred pages and fifteen hundred commits, nearly all of them made by Codex or Claude. It's the single most useful thing I've built for working with agents, and the reason is boring. It's files.
Personal tooling, used daily, shared with agentsI spend all day, every day, prompting agents. Not writing code, and not writing one big prompt either. On a normal weekday I have somewhere between eight and ten Codex threads live at once, a couple of Claude Code sessions beside them, and a few dozen automations ticking in the background. Last Monday Codex opened thirty-five threads for me. Most of them I never typed into, because they were started by another thread, or by a timer.
How I work day to day, from a month of transcriptsI've tried most to-do apps. The problem was never the app. It was that the things I commit to come out of stand-ups, Slack threads and agent sessions, they get extracted by Claude or Codex into a scratch window, and then the window closes. There was no one place that said what I'd personally promised, how it mapped to my open PRs, and what I owed somebody by when.
Personal tool in daily useA relative was building a house in rural County Cork. The council had published the whole planning application, drawings and all, so I wanted to know whether an agent could turn those scanned sheets into a model I could stand inside, and whether I could trust it.
Side experiment, two days in September 2026All real, all shipped. Mountain rescue, maps, music, a house, a to-do map an agent can see.
Mission control for mountain rescue
Live GPS tracking and mission management for an Irish mountain rescue team, built to replace a seventy-thousand-line QGIS plugin. The installed Electron app is the real runtime: SQLite, offline map packages, recovery on restart. Final acceptance as the plugin's replacement is still open. The browser version is a team-testing path that keeps data in the session only.
A public planning application, modelled and checked by agents
Scanned council drawings turned into a scripted Blender model, three renders and a fifteen-kilobyte WebGL walkthrough. The verification loops were the work.
A to-do map the agent can see
A spatial canvas of tasks and ideas that Codex drives through WebMCP page tools. Reads tasks live, owns only presentation, refuses stale reviews. Local only; shown here with fictional data.
A step sequencer with patterns and grooves
Five channels, sixteen steps, four pattern slots, groove templates and a procedural Web Audio kit. No framework, no build step. The webcam hand tracking from the first version came out in the October redesign.
Six map series of Ireland, 1829 to 1943
Overlay, swipe, side-by-side and lens comparisons across six historic series, with twenty-seven live public archaeology and heritage sources on top.
A weekly brief for an Irish farm
Pin a farm and get the week: rain and gust windows, CAP and LPIS parcels, OPW water levels, EPA data, market prices. All public Irish data, no login. Its methodology page says what is stale, incomplete or differently scoped, which I think is the interesting part.
Updated monthly.
Working on: the harness around a production agent platform: which tools an agent sees first, how it asks fewer unnecessary questions, how its work is proven before a PR.
Thinking about: how much process a team needs when the models are this good, and how to tell when a rule has stopped earning its place.
Building on the side: a to-do map that Codex can see and talk about at the same time, with a small model organising it in the background.
I build and operate production AI agent platforms at Alludium. Before that I was a Principal Software Engineer and AI Champion at Dell, where I drove AI adoption across engineering teams. I live in Cork.
Most of my day is spent on the system around the model: context, harness, memory, validation and orchestration. On the side I ship things for Irish mountain rescue, Irish public data, and whatever I want to understand next.