One day last July, an AI coding agent working for the American investor Jason Lemkin did something no employee would survive. Under an explicit code freeze — no changes, the instruction could not have been plainer — it deleted the live production database, invented data to paper over the hole, and reported that the loss was irreversible. It was not; Lemkin recovered the records himself. The tool was Replit, and Lemkin was doing what 2025 taught the world to call vibe coding: prompt, accept everything unread, ship, and — in the words of the man who coined the phrase — “forget that the code even exists”.
Hold that scene; we will need it later. But it is not what this piece is about. The story that matters to a financial services firm is the quieter one that was unfolding at the same time inside the world’s most sophisticated companies — and what it now demands of the firms that follow them.
From vibes to agents
Andrej Karpathy — a founding member of OpenAI and one of the field’s most respected voices — coined “vibe coding” in February 2025 in a half-joking post he later dismissed as a shower thought, and was candid that he meant it for throwaway weekend projects. Within weeks the term had escaped him: as the programmer Simon Willison warned at the time, it was soon being applied to every kind of AI-assisted development, including the careful, professionally supervised kind that is its exact opposite. By November, Collins Dictionary had crowned it word of the year. Within the year, its own inventor had pronounced it passé — and reached for a better name for what serious practitioners were actually doing: agentic engineering.
The distinction is everything. In vibe coding, nobody reads the output and nobody checks. In agentic engineering, AI agents plan, write, test and iterate across an entire codebase — doing the typing, and increasingly most of it — while a professional specifies the intent, sets the gates, reviews what matters and owns what ships. One is a toy with a blast radius. The other is the largest change in how software is made in a generation, and it is the subject of this piece.
The numbers are in
The evidence no longer rests on demonstrations. Microsoft says 20 to 30 per cent of the code in its repositories is now written by AI; Google puts its own figure above 30 per cent. A quarter of Y Combinator’s recent intake arrived with codebases that were 95 per cent machine-generated, and Anthropic’s chief executive predicted last year that AI would shortly be writing 90 per cent of all code. Discount the measurement games and the salesmanship as you please; the direction is not in doubt. Nor is finance watching from the stands. JPMorgan has reported engineering productivity gains of up to 20 per cent from its AI assistant. Goldman Sachs has gone further, setting Cognition’s Devin — an autonomous “AI software engineer” — to work across its twelve-thousand-strong developer force, under human supervision, on legacy modernisation and debugging. The commercial scale tells the same story: Cursor, the professionals’ agentic editor, went from a standing start to $2bn of annualised revenue in about three years, the fastest ramp enterprise software has recorded; Lovable of Stockholm reached $400mn inside two; Microsoft’s Copilot sits on millions of desks earning an estimated billion dollars a year. Whether the valuations riding on all this hold is a question this series has weighed before. That the capability is real is no longer seriously contested.
One consequence arrives before any tool is chosen. Change in this industry is substantially delivered by systems integrators — TCS, Infosys, Accenture, Cognizant, Capgemini — whose model is the sale of hours. A technology that lets a tenth as many people produce the same output is not an efficiency for that model; it is an existential threat to it. The contract negotiation of the decade is the move from pricing the body to pricing the outcome, and the supplier’s every incentive is to resist, because the hours are the product. Having spent years on the buying side of those contracts, I would say only this: rewrite them before your suppliers are ready, not after.
Gains, with conditions
So the answer to “will this make change cheaper and faster?” is yes — genuinely, materially yes. But the gains are conditional, and the conditions are where a finance leader earns their keep. The most instructive study of last year, a controlled trial by the research group METR, timed experienced developers working on large, mature codebases they knew intimately, with and without AI. They believed the tools had made them about 20 per cent faster; measured, they were 19 per cent slower. That is not evidence the technology fails. It is evidence that the gains depend on where and how it is deployed: spectacular on greenfield work, treacherous when pointed naively at a forty-year-old estate of the kind anatomised at length elsewhere on this site. Practitioners call the pattern the seventy per cent problem — the agent races to a rough three-quarters of a solution and bogs down in the hard last quarter. Goldman’s own chief executive put it perfectly, describing AI drafting the bulk of a prospectus in minutes: the last 5 per cent is what now matters, because the rest has become a commodity.
Read that way, the prize for an established institution is specific and large. The long tail of small enhancements that today die unfunded in the prioritisation queue becomes affordable. And the richest, least-advertised gain is the inverse of the start-up narrative: not writing the new but comprehending the old — the undocumented COBOL that three engineers near retirement still understand, the tangle nobody dares touch. It is precisely there that Goldman aimed Devin. The greatest service AI may render a financial institution is not to write its future but to explain its past.
Accountability, rebuilt
All of the above will shortly be available to everyone, which is why none of it is the differentiator. The differentiator is the thing the Replit story was actually about. The agent’s sin was not bad code; it was unauthorised action and a falsified record — a broken change freeze and a misreported outcome. And here is the uncomfortable truth: our entire accountability model for software assumed that humans wrote it. Responsibility attached to authorship — the developer who wrote the lines, the reviewer who read them, the change board that blessed them. When agents do the typing at machine speed, that model does not merely creak; it misses the point.
What replaces it is accountability by authorisation. The accountable human no longer vouches for having written or read every line; they govern what the agents are permitted to touch, define the gates every change must pass, and answer for the framework and the release. Review itself becomes substantially machine-on-machine — one agent writes, a second critiques, a third generates the adversarial tests — and should, because that is the only kind of review that scales to machine-speed output. But a consensus of models is a quality mechanism, not an accountability mechanism: three agents agreeing is admissible evidence in the file; it cannot be the signature. Beneath the machine consensus sit the harder guarantees — verification by execution, meaning behaviour proven under test rather than text that merely reads well, and the complete, immutable log of every agent action, which becomes the new audit evidence. Regulators, as it happens, already speak this language. DORA has been in force since January 2025; the UK’s operational-resilience regime and the senior managers’ framework want a named, personally accountable human. The name on the accountability map does not change. The evidence file behind the name changes radically — and building that file is the real project.
The skills consequence follows directly and deserves board attention now. Value migrates from writing code to specifying and judging it, which hollows out the junior rung on which every engineering career was built — Bloomberg Intelligence already projects up to 200,000 banking jobs going over the coming years, entry-level first. If juniors no longer learn the craft by writing, who is qualified to supervise the machines in ten years? A firm that does not deliberately rebuild that apprenticeship is eating its seed corn.
The signature survives
Is agentic engineering a positive for financial services? Unambiguously — faster and cheaper with it, in the places and on the terms set out above. But the firms that convert it into advantage will be the ones that treat it as what it is: a redesign of the operating model and the control framework — kin to the function-by-function discipline the previous piece in this series applied to premises — rather than a procurement decision. The machines will do the typing; increasingly, they will do the checking too. The signature stays human. Everything behind the signature is what changes.