[What you will be working on]
The AI Programme builds shared AI infrastructure that agencies can run in production, instead of rebuilding the same foundations independently. It sits within Government Digital Products, and it owns a small number of central products that a lot of other people’s work now depends on.
Three of them shape the job. There’s the agent surface public officers use to get through the day: it reaches into their mail, their documents and their case systems, so it has to be quick, honest about what it just did on their behalf, and safe to point at real records. Underneath sits the developer platform, where teams across government bring their own agents, MCP servers and skills and move them into production through review, approval and audit. And there’s the model and capability layer that most agencies meet first: a compliant gateway to multiple production-approved models behind an OpenAI-compatible API, alongside shared speech, vision and document capabilities (live transcription, image and video classification, key information extraction) that every agency would otherwise go and procure separately, with guardrails, quotas and cost attribution already attached.
Building an agent demo is easy now, and that’s fine, because it means the hard parts have moved. Most of what we work on lives after the demo. An agent decides to call a tool that sends an email on someone’s behalf, and the approval gate for that has to be enforced by the runtime, not requested politely in a system prompt. A tool loop needs a stop condition that doesn’t depend on the model deciding it’s done. Token spend has to trace back to the team that caused it, and when an agency’s agent gets slower this week, someone has to be able to say why. Rollback gets interesting when the artefact is a prompt, a tool schema and a model version rather than a binary. Very little of that is model work. It’s orchestration, platform and operations, which is roughly the job.
You’ll work across all of it: the surfaces officers see, the runtime underneath, the platform other engineers build on, and the unglamorous business of keeping the whole thing up. Your user is a public officer with something to finish today. Your customer is often another engineering team in another agency.