If your team already works with coding agents, the complaint may sound familiar: the day goes on reviewing what the agent writes. Anthropic has just put numbers on it inside its own company, and the conclusion is not that the agent is slow. It is that the agent waits for a person.
The idea this insight argues: meta-work (preparing what you ask the agent, reviewing what it returns and deciding what goes in) is not a distraction from the real work. It is the work. A team that does it in odd moments and without a method turns the agent into a queue of things to review. What cuts that waiting is deciding upfront, not reviewing faster.
What Anthropic measured in its own company
On 16 October 2026, Anthropic's economics team (Wilmers, Catipon, Massenkoff and McCrory) published how work has changed inside the company since its employees started using agents. Three figures explain the problem:
The 80% is calculated by subtracting the agent's active time from the total time of each interactive session. By meta-work they mean the work that is not the task itself: writing specs for multi-agent runs, having one agent review another's output, building tools to see what is going on and monitoring pipelines. And their conclusion: this acceleration makes human attention the bottleneck for agent output.
Risk It is a single company, the one that builds the model, studying its own deployment. The 80% is session wall-clock time and includes short waits: it does not mean the agent sits idle 80% of the working day. Meta-work was classified automatically on a sample of 2,000 one-line summaries. Six agent hours per human hour is not productivity: the authors themselves warn that more code does not necessarily mean better code or business impact.
Why the agent waits
It is not an isolated figure. In December 2025, another Anthropic study of its engineers (Huang et al.; 132 surveyed and 53 interviewed) already pointed to the cause. Engineers mostly delegated what is easy to check: they said Claude was very useful where validation effort is not large compared with creation effort. And more than half said they could fully delegate only between 0% and 20% of their work. They called it the "paradox of supervision": supervising the agent requires the very skills that may weaken from overusing it.
DORA frames it as a cost. Its report on the return of AI in software development, as summarised by InfoQ (May 2026), speaks of the "verification tax" of reviewing AI-generated code, and warns that more code moving faster can overwhelm deployment pipelines and manual review. And its 2025 report sums it up in one sentence: AI amplifies what the organisation already has, good and bad.
Putting it together, our reading: the agent waits because what it returns cannot be taken as good without someone looking at it, and the person who looks arrives late because they did not know what to check. We already covered where the gain gets stuck in more code, same delivery; here the bottleneck is wider: not only in review, but across the whole flow.
Meta-work is the work
If a person's attention is the scarce resource, the sensible thing is to spend it where it pays most: before launching the agent, not after. The table compares the three parts of meta-work done in odd moments and done with a method. It is our proposal, not a finding of the study.
| Part | Done in odd moments | Done with a method |
|---|---|---|
| Prepare | One line in the issue and the agent fills in the rest | A spec with scope, limits and acceptance criteria that can be checked |
| Review | Whoever is free reads the whole diff, line by line | A machine checks what is checkable (tests, analysis, executable criteria) and the person looks at the rest |
| Decide | Decisions happen when the agent asks or when something breaks | The points where a person decides, because there is risk, are set before starting |
Prepared by onext
The three cells on the right are what Spec-Driven Development (SDD), building software from specifications, does before the agent writes a single line. Preparing means turning the task into a spec, as we explain in an issue is not a spec. Reviewing starts by turning acceptance criteria into the merge check. And deciding means knowing upfront where you sign off and how to undo what the agent does.
Every decision taken before launching the agent is one wait less afterwards. A person decides where there is risk; everywhere else, what is already written decides.
What changes in the team's planning
Two concrete changes, which we propose:
- Meta-work gets planned. Writing specs, acceptance criteria and the team's rules for agents counts in the sprint as work, with its own time. If it is done in the gaps left by delivery, it is done badly or not at all, and the agent waits again.
- Review is designed by risk. It is not handed to whoever is free. A text change on a screen and a change in how an invoice is calculated do not deserve the same person or the same time. That is decided beforehand and written down.
An example, made up and from no client: a team asks the agent to "export orders to CSV". Without preparation, the agent asks which columns, which date format and what happens with cancelled orders, and waits; or it does not ask and decides itself, and the reviewer finds out in the diff. With half an hour of preparation, the spec already states the columns, the format, that cancelled orders are left out and that a test checks it. The reviewer only has to look at what the test does not cover.
This time spent preparing and reviewing is exactly what new people need to be taught, as we explain in junior developers: learning to review: if supervising requires the skills the agent can erode, they have to be practised on purpose.
What this figure does not tell you
It does not tell you how long the agent waits in your team. Anthropic is the company that builds the model, with a team that works with it every day and conditions that are not yours. Nor does it say that more agent hours mean more delivery: what you measure to know whether it works is delivery and its quality, not how much code comes out or how many hours the agent works.
What it does say, and what we keep, is the direction: when the agent does more, the limit becomes the attention of whoever prepares, reviews and decides. And that can be organised.
If you want to see, with someone from outside, where the agent waits in your team today, the diagnostic is free, and if the diagnostic shows no difference, there is no Quick-start. The rest of the path is at AI-Accelerated Development. The full method is in Spec-Driven Development with AI.
Frequently asked questions
What is meta-work when a team works with AI agents?
It is the work that is not the task itself but what prepares and watches it: writing specs for agents to work from, having one agent review another's output, building tools to see what is going on and monitoring pipelines. Anthropic measures it in its own company and found it rose from 10% of interactive Claude sessions in fall 2025 to 22% in spring 2026.
Does a coding agent really spend 80% of the time waiting?
In the interactive sessions Anthropic measures in its own company, yes: around 80% of a session's wall-clock time the agent is waiting for a person to respond. It is calculated by subtracting the agent's active time from the session's total time, so it includes short waits. It does not mean the agent sits idle 80% of the working day, nor that it is your team's number.
Why is a person's attention the bottleneck?
Because the agent produces faster than someone can read, check and decide. Anthropic puts it this way: the acceleration makes human attention the bottleneck for agent output. DORA calls it the "verification tax": the cost of reviewing AI-generated code.
How do you reduce the time the agent waits for a person?
By deciding upfront, not by reviewing faster. Before launching the agent: what is being asked and with which acceptance criteria, what a machine checks without anyone looking (tests, analysis, executable criteria) and at which points a person decides because there is risk. Whatever is left open comes back as a question or a doubtful diff.
Is meta-work wasted time that should be cut?
No. It is the part of the work that decides whether what the agent does is any good. What is a problem is doing it in odd moments, without a method and stealing time from delivery. This insight proposes planning it as work: writing specs and team rules counts in the sprint just like writing code.
What are the limitations of the Anthropic study?
It is a single company, the one that builds the model, studying its own deployment. Meta-work was classified automatically on a sample of 2,000 one-line summaries, tool adoption is not random and there are 17 interviews. The authors themselves warn that more code does not necessarily mean better code or business impact.
Sources
- Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory (Anthropic), "How employment is changing at Anthropic", 16 October 2026.
- Saffron Huang et al. (Anthropic), "How AI is transforming work at Anthropic", 2 December 2025.
- Matt Saunders, "New DORA Report Claims Strong Engineering Foundations Drive AI Return on Investment", InfoQ, 11 May 2026 (on DORA's "ROI of AI-Assisted Software Development" report).
- DORA, "State of AI-assisted Software Development 2025", 2025.
- The table, the two planning changes, the example and the reading of why the agent waits are onext's own work.
Bernat López is founder and CEO of onext, an AI boutique. He helps development and product teams work with AI with method —specification before building, a person who decides where there is risk and Spec-Driven Development— and applies to his own company what he proposes: onext runs on its own agentic system.
LinkedIn →