We build with agentic AI every day at Shape. We're also reining it in.
That probably sounds contradictory for a software agency with global clients and an AI pipeline that integrates with our Kanban board and drafts pull requests while we sleep. But it's where eighteen months of using these tools in production has landed us, and it's a position the wider UK technology sector needs to take seriously.
Our agent pipeline reads a feature ticket, scopes the work against the relevant code, and produces a draft pull request via Cursor's background agents. By the time an engineer sits down, half the work is done. The first time it shipped a clean PR was genuinely exciting.
The potential here is exciting and transformative. In legal services, we'll see agents efficiently gathering case precedents and crafting compelling opening positions. In healthcare administration, coordination agents will expertly manage referrals. In manufacturing, maintenance agents will strategically schedule parts orders before failures occur. Regardless of the field, we're tapping into universal principles like planning, tool use, memory, and coordination, setting the stage for incredible advancements.
Yes, it all sounds amazingly efficient and is a business owner's dream. Here's the part that tends to get skipped. Two things have become uncomfortably clear to my co-founder and I after a few months.
First, AI gets it wrong. Frequently. And not in spectacularly obvious ways. It makes small, plausible, very convincing judgments that pass a quick read and only surface later when something behaves oddly in production. Spotting these errors is harder than it sounds, because plausibility is not the same as correctness, especially if you don't have the experience to know what correct looks like or what trade-offs you're actually making.
Second, our engineers are getting weaker. Not all of them, not all the time, but the pattern is real. Juniors who never have to wrestle with a problem on their own are losing the intuition that lets them spot when an agent's solution is subtly off. I noticed it in myself when I sat down recently to write some code without AI help and struggled with patterns I had shipped to production the week before. The code I'd written previously was technically owned by me. But it wasn't necessarily owned by my brain.
So we're reconfiguring our pipeline. Instead of letting agents do the interesting first-draft thinking, we're narrowing them to the busywork: boilerplate, scaffolding, repetitive plumbing. The problem-solving stays with the engineer. Faster than working unaided, but the thinking stays human.
Most of the UK leadership discourse around AI measures success by adoption rate: how many companies are using it, how deeply it's embedded, how fast the productivity numbers move. That's the wrong metric.
In twenty years, the senior engineers, lawyers, and healthcare professionals running the UK industry today will be retired. The people replacing them are the juniors entering those professions now, the same juniors, if they're lucky enough to get hired, being told to "just use AI" and that "AI is better than you anyway." If they spend the formative decade of their career blindly following AI output, we won't have a generation of senior practitioners to take the wheel. We'll have a generation of supervisors who never built the underlying competence.
That isn't a culture of excellence. It's a culture of good enough, because it got done quickly, with minimal effort, and nobody had to think too hard about whether it was actually right.
The obvious counter is that people said the same thing when Google launched, and they were wrong then. But Google was a different paradigm. Search gave you raw materials and required you to do the work: read sources, evaluate them, synthesise, and apply. The cognitive load shifted; it did not disappear. Agentic AI produces the synthesised output. It writes the code, drafts the contract, and presents a finished artefact for approval. The work is to evaluate, not to build. And evaluation without the experience of building is precisely how you miss the things that matter.
If the UK wants to lead on AI in a way that actually compounds, the measure should be the strength of our practitioners in twenty years, not the speed of adoption this quarter.
The shifts UK tech needs to make, on our reading, fall under three headings.
Technical. Build agents that show they are working. Traceable plans, explicit tool calls, and human-reviewable diffs are what make it possible for the person in the loop to learn from the work rather than just rubber-stamp it. We've been thinking about a "learn mode" for our own pipeline: one that walks the engineer through why each decision was made and invites them to critique it. Adoption shouldn't have to mean passivity.
Organisational. Treat agent output as a first draft to interrogate, not a result to consume. That means designing in pair time, deeper code review, and deliberate practice without the agent. Otherwise, you ship faster and atrophy quietly.
Policy. Apprenticeship standards, skills frameworks, degrees and accreditation routes need to assume the new baseline: working effectively with agents and being able to work without them. A generation trained only as agent supervisors, with no foundation underneath, is the worst outcome our sector could produce.
This is part of why we're a founding member of Agents.ne, a not-for-profit that helps junior engineers navigate the narratives aimed at them today. "You may never get a job." "Just use AI." "AI is better than you anyway." None of that is true, and all of it is corrosive: to careers, to the talent pipeline, and to the long-term capability of UK tech.
The juniors coming through now are the seniors we'll depend on in five years. Agentic AI will be part of how they work. Still, it cannot be a substitute for the experience that produces real engineering judgement. Our job, as employers, as a sector, and as a community, is to make sure they get both.
The technology will keep getting better. Whether our people do is up to us.
