.webp)
Last quarter, one of our senior engineers showed me a feature he had finished in an afternoon. A year ago it would have taken his squad most of a sprint. I did not feel excitement first. I felt a quiet question: If this is normal now, what exactly am I paying people to do?
I have spent the month since working out the answer, and the 2026 Agentic Coding Trends Report gave me language for it. I do not agree with every prediction in it. But the central point matches what I see in my own company.
The number that changed how I think
Anthropic's research found that developers use AI in roughly 60% of their work, yet they say they can fully delegate only 0 to 20% of tasks.
That gap is the story. Most commentary on AI and engineering swings between "it replaces everyone" and "it is just autocomplete". The data says neither. AI is a constant collaborator that still needs a person watching it.
For a CEO, this matters more than any benchmark. If your plan assumes headcount disappears, you will be disappointed. If your plan assumes nothing changes, you will be overtaken.
From writing code to directing it
The report argues that engineers are moving from implementer to orchestrator. I agree, and it shows up in how I now judge talent.
I used to ask how fast someone could build something. Now I ask how well they can break a problem into pieces, hand those pieces to agents, and tell whether the result is right. Those are different skills. The second set looks a lot like management, and many brilliant coders have never practised it.
One engineer in the report put it well: He uses AI where he already knows what the answer should look like, and he built that instinct by doing the work the hard way. I think about that line often. It means we still need to train juniors properly, even when an agent could do their first-year tasks. Skip that, and in five years we will have nobody who can tell good output from plausible output.
From one agent to a coordinated team
The report's second trend deserves attention: Single agents giving way to coordinated teams. An orchestrator splits a job into parts, hands each to a specialist working in parallel with its own context, then combines the results.
Fountain, a workforce platform, did this for hiring. Screening ran 50% faster, onboarding 40% quicker, and one logistics customer staffed a new fulfilment centre in under 72 hours, down from a week or more.
That is a management problem more than a technical one: Break down the work, assign roles, define handoffs. Our tooling also assumes human pace, so the groundwork starts now.
Review is the new bottleneck
This is what keeps me up. When agents produce more code, humans do not get more time to review it. The volume goes up, and our attention stays the same.
The report predicts that agents will review other agents' output, and that people will focus on what matters: Strategic decisions, unusual edge cases, anything touching the business. I think that is correct, but it needs deliberate design. You cannot just add a second AI and hope.
So the question I now ask every team lead is simple. Where does a human look, and why there? If the answer is "everywhere", we are not scaling. If the answer is "nowhere in particular", we are gambling.
For a company in quality engineering, this is both a threat and an opportunity. Testing and verification stop being a final gate and become the main way humans stay in control. The firms that build trustworthy checking systems will matter far more than the ones that only generate code faster.
More output, not just faster output
The report includes a detail I found more interesting than any headline. About 27% of AI-assisted work, it says, consists of tasks that would not have been done at all otherwise. Think of the small annoyances everyone deprioritises, or the internal dashboard nobody had time to build.
That reframes the productivity conversation. I stopped asking "how much faster can we ship the same roadmap?" and started asking "what has been sitting on the backlog for two years because it was never worth an engineer's week?" Often those neglected items are where customers feel the most friction.
Software is no longer only for engineers
Another trend worth taking seriously is that people outside engineering are starting to build their own tools. The report describes Anthropic's own legal team, where a lawyer with no coding background built self-service workflows that cut review turnaround from days to 24 hours.
I have started encouraging this in our own company. Operations, sales, and finance each have repetitive work that never earns a ticket in the engineering queue. If the people closest to the problem can fix it themselves, everyone wins. The risk is obvious, though. Tools built without engineering discipline can leak data or break quietly. So we are pairing enthusiasm with simple guardrails: Approved data sources, a short review for anything touching customer information, and a named owner for every tool.
Security cuts both ways
I will be blunt here. The same capability that lets any engineer run a decent security review also lets attackers scale their work. The report says this plainly, and I would rather hear it plainly than wrapped in optimism.
My response is to make security part of the design from day one, not a cleanup job before launch. It costs less early and a great deal more later.
The honest uncertainty
I do not know how fast this moves. The report itself says its predictions reflect what customers are doing today, not certainties about tomorrow, and I appreciate that candour. Some of these trends will arrive slower than advertised. A few will surprise us.
What I am confident about is the direction. Writing code is becoming the easy part. Deciding what is worth building, checking that it works, and keeping people accountable for the outcome: That is where the value goes.
Companies that treat this as a passing productivity tool will find themselves playing a different game from everyone else. I would rather be early and slightly wrong than late and certain.
The goal was never to remove people from the loop. It is to make sure their attention lands where it counts.
.webp)
