For about a year I had something almost nobody in Australia got: permission.
Not a budget line, though there was one of those, and not a job title, though there was one of those too. Permission. The genuinely rare kind — to go deep and wide, to play with and break and rebuild, to treat a regulated health insurer's software development lifecycle as a thing I was allowed to take apart on the workbench and put back together better. Mathew Finch and Brendan Mills, who led Data & AI and technology respectively, looked at the wave coming and decided the right move wasn't a steering committee. It was to hand someone the keys to the candy store and see what happened.
Reader, I binged on the sweet sweet sugar. Twelve, fourteen, sixteen-hour days. The kind of obsessive, joyful, slightly-unhinged building you only do when someone you respect has effectively said "go find out." I cast a vision for what an AI-centric SDLC could actually look like — not the LinkedIn-keynote version, the real one, the one that has to survive cyber review and a governance posture — and then I went and built it. The result got a name with rather a lot of hubris baked in: The Prometheus Project. Steal the fire, hand it to the engineers, try not to get my liver pecked out.
It started as eleven tools. Thirty-to-forty grand a month in tokens later it was eighteen.
I'll write each of these up properly in their own time, but the shape of it: a Governor Module that turned policy into code you inject at invocation and assert on in CI, so cyber and design and cloud could each have a seat at the table without sitting in every meeting — governance as a team sport instead of a tollbooth. Skynet, a context graph and code-intelligence layer spanning all 3,600-plus repos in the org, giving coding agents real upstream, downstream and cross-repo sight instead of the goldfish memory they ship with. JEFF, a cloud-based multi-agent coder you could summon from Jira, Slack, Claude Code or the API — and yes, in Jeff Goldblum's voice, because if you can you must. Echo, a seven-agent PR-review army modelled on Anthropic's multi-agent work, wired into our standards and cyber and governance, with a learning loop so it slowly worked out what it was and wasn't allowed to wave through. An auth broker, an AI gateway, a cyber platform for MCP and skill security, Dojo for training the humans, and Signal, a developer-experience instrument ingesting fifty-eight datapoints to tell us — honestly — whether any of this was actually landing.
That last one matters, because here's the part that surprised me.
The tools weren't the hard bit. The hard bit was people, and not in the way the headlines mean it.
When you give engineers agentic tooling, the coding gets fast. Stupid fast. The outliers — already exceptional engineers — were moving end-to-end at two-to-four times their old pace, one person doing the work of four. But most people landed around a thirty-to-forty percent bump, and overall team productivity in some places actually went down, which looks like a paradox until you do the arithmetic. Coding is about sixteen percent of an engineer's day. Pour rocket fuel on that sixteen percent and you don't shrink it — you just generate three or four times the volume of decisions, PRs, reviews, releases, QA and support that make up the other eighty-four. You get a close-but-not-quite mythical man-month problem, because AI is broadly good at coding and not specifically good at being an engineer in your org.
Most companies are speeding up the sixteen percent, watching the eighty-four percent jam, and concluding that AI is broken. It isn't. Their plan is. They've quietly turned themselves into companies that manufacture bottlenecks instead of software. A good chunk of my year was spent building the counterweight — the tools and the practices that accelerate the other eighty-four percent to match — which is the whole reason Prometheus is a platform and not just a fancy IDE.
But even the bottleneck wasn't the real lesson. The real lesson is older and less comfortable: the degree to which a company gets accelerated by AI has almost nothing to do with the technology and almost everything to do with its people. AI is an accelerant. It speeds you toward whatever outcome you were already barrelling toward — good, bad or mediocre — because of the assumptions you already hold. It will gladly take your terrible idea, tell you it's fantastic, and help you polish that turd faster than you ever dreamed. For someone else, on a good day, it's an alchemist turning lead into gold — but that person was usually a decent chemist to begin with.
The headwinds in nearly every stalled AI project trace back to two old ghosts holding hands: the Peter Principle and Conway's Law. The wrong people leading for the moment, and an inability — or an unwillingness, or a lack of trust — to communicate across boundaries. People with a headline understanding of AI making calls about work done by people with a deepening, hands-on understanding of it. The cure isn't a tool. It's curious humility, a willingness to listen, and an actual appetite to experiment.
Which is the segue I owe to Emma, who took over leading Data & AI partway through. She did the thing very few leaders manage: she made me care about governance instead of begrudgingly tolerating it. Under her, governance stopped being the thing that slowed the fire down and became part of how the fire was allowed to spread safely. Her fingerprints are on everything I built — quite literally, because she's the reason every tool I ship has a light mode, a dark mode, and a pink mode. If it doesn't have pink mode, it doesn't ship. That's not a quirk; that's a leader teaching you that the people using the thing are the point, and you don't get to decide for them what "usable" means.
Did it work? The people who actually adopted the tools and the practices moved at that two-to-four-times clip and kept it. We ran hackathons, workshops, AMAs. We pressure-tested ideas against teams at Endava, Mantel, Google, Cursor and AWS, all of whom were genuinely generous with their time. I got to work shoulder-to-shoulder with some of the sharpest humans I've shared an org with — Emma, Kurt, Brad, Pip, Ben and a long list of others who made the candy store feel less like a solo sugar binge and more like a team sport.
So that's the year. The rare freedom, the eighteen tools, the fifty-eight signals, the liver mostly intact.
If there's one thing I'd hand to anyone trying to do this inside a real company — bank, insurer, anything with a compliance team and a risk appetite — it's this: don't buy the tools first. Buy the permission. Find the leaders with the curious humility to say "go find out," protect the people doing the finding, and accelerate the whole lifecycle instead of just the bit that demos well. The technology is the easy part now. It always was the people.
I'm leaving the candy store to go build at the edge of all this, which means it's time to go full Gilfoyle, update the profile, and wait patiently by the door for the recruiters and their loot bags. That happens in real life, right?
Thank you, nib. It was a hell of a fire.
Stay curious. Stay humble. Keep experimenting.