What’s trending in AI on 17 October 2026: Most claims about AI and jobs are forecasts. This week one of the companies building the technology published something rarer: its own payroll-level data. In a study released on 16 October, Anthropic’s economists report that AI agents now log about six hours of work for every hour a human spends directing them, that the junior share of the company’s technical hires has fallen from 27% to 9% since early 2025, and that engineers’ code output per person has roughly quintupled in six months. The same report names the catch: about 80% of an interactive agent session is spent waiting for a person to respond. Human attention, not compute, is now the limiting factor. A day earlier, PostgreSQL showed what that looks like in the wild, when AI code reviewers found more bugs than volunteer maintainers could fix and the release of Postgres 19 slipped by about a month. Below: what Anthropic measured, what it means for junior roles and managers, why review capacity is the new constraint, and an eight-question audit for your own team.
Key takeaways
- Agents out-work people, by hours. At Anthropic, agent work hours overtook human hours in March 2026, and by August agents were working six hours for every human hour in employee-directed sessions.
- The entry rung is shrinking. Junior staff fell from 21% to 14% of technical employees, and from 27% to 9% of technical hires. About a third of ML managers said they were hiring less because of AI.
- Output is up, quality held. Code commits per engineer reached about 5x their level six months earlier, and agent-written code was no more likely to be deleted or reverted than the same engineer’s own code.
- People are the bottleneck. Interactive sessions are about 80% idle time waiting on a human; meetings rose 20%, and work spent managing agents more than doubled.
- Review is the new scarce skill. Postgres 19 was delayed to around 19 November after AI review surfaced dozens of small defects and 8 to 10 features were pulled. Plan for checking, not just generating.
1. What Anthropic published, and why it matters
The study, titled “Employment at Anthropic” and written by economists Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory, treats the company as an early test case: a workplace where frontier AI tools arrived before they reached everyone else. The authors drew on internal headcount and financial records, employee surveys, calendar and communication data, hiring scorecards, telemetry from agent sessions and 17 interviews.
Two caveats belong at the top. First, Anthropic sells the tools it is measuring, so a report showing those tools make staff more productive is not neutral evidence; read it as a well-documented case study, not an independent audit. Second, the authors flag their own main limitation: employees chose whether to adopt coding agents, and people working on big projects may have adopted first, which could flatter the productivity figures. They try to control for this by comparing each engineer with themselves over time, but it is not a randomised trial.
Even with those caveats, it is the most detailed public look yet at what happens inside an organisation when agents are given real work at scale. It also lines up with what we found in our piece on why individual AI gains rarely show up at team level: the gain is real, but it piles up wherever humans have to approve, review or decide.
2. The numbers: agents, output and hiring
The headline figure is about hours. In sessions where employees direct agents, agent work time passed human work time in March 2026. By August, the ratio was six agent hours to one human hour. Output followed: code commits per engineer reached about five times their level of six months earlier, and the company’s revenue per employee at around five years old was two to four times what Meta and Google managed at the same age, according to the authors’ comparison of inflation-adjusted 10-K figures.
The productivity check is more careful than a before-and-after chart. The team followed 192 engineers who took up coding agents and 74 who did not, and saw output rise sharply after adoption. On quality, the study compared agent-written and human-written code from the same person in the same month and found no difference in how often it was later deleted or reverted. Non-engineers changed too: coding-focused agent sessions run by non-technical staff went from about 2 a month in August 2025 to about 90 a month in May 2026. One account executive estimated that automation took over about a fifth of their old tasks and let them talk to 30% more customers.
| Measure | Earlier | Latest |
|---|---|---|
| Agent vs human hours (employee-directed sessions) | Agents behind humans until March 2026 | 6 agent hours per human hour (Aug 2026) |
| Code commits per engineer | Baseline six months earlier | About 5x |
| Junior share of technical staff | 21% (early 2025) | 14% |
| Junior share of technical hires | 27% (early 2025) | 9% |
| Technical departments’ share of headcount | 61% | 56% (a year later) |
| Agent-management (“meta-work”) share of interactive sessions | 10% (fall 2025) | 22% (spring 2026) |
| Standing instructions written for agents | A few hundred (July 2025) | About 8,500 (a year later) |
| Staff agreeing their processes make them as productive as possible | 44% (late 2024) | 75% (April 2026, question reworded) |
3. The junior hiring squeeze
The finding that will travel furthest is about who gets hired. Junior employees made up 21% of Anthropic’s technical staff in early 2025 and 14% now. Among new technical hires the drop is steeper, from 27% to 9%, and postings for junior roles fell sharply after December 2025. In an April 2026 survey, a majority of managers said AI had changed the skills they hire for; about a third of machine-learning managers and a fifth of software engineering managers said they were hiring less because of it.
What managers look for has shifted too. Interview scorecards now mention candidates’ use of AI tools three to four times as often as in 2024. For technical candidates, comments about autonomy and judgment grew most, while comments on domain expertise declined. For non-technical candidates, mentions of customer and mission focus rose while execution skills fell. In plain terms: if an agent can do the execution, the person is hired for deciding what to execute and whether the result is right.
The authors do not pretend this is fine. They note that internal protections, such as helping existing staff move to new roles, do nothing for people who were never hired, and they call for public policy on retraining and income support. For other employers, the uncomfortable question is a pipeline one: if you stop hiring juniors, where do your future seniors come from? Judgment is learned by doing the execution work that agents now absorb. Our coverage of Berkeley’s study on AI and persistence found a related risk: people who lean on AI for answers rather than hints can lose the habit of working through problems.
4. The new bottleneck is you
The most useful finding for other businesses is less dramatic than the hiring data. Interactive Claude sessions are roughly 80% idle time, spent waiting for a person to read, decide or reply. For non-technical teams, simply reading everything their agents produce would take more than half a working week. The agents are not slow; the humans supervising them are busy.
The organisation is adapting around that constraint. The share of interactive sessions spent on “meta-work”, such as building tools for agents, monitoring pipelines and writing specifications, rose from 10% to 22% in about six months. Standing instructions written for agents grew from a few hundred to about 8,500 in a year. Median meeting time rose 20% year on year, and 16% for technical individual contributors, presumably because coordination is now the work. The engineer-to-product-manager ratio is about 17 to 1, and managers’ own code commits doubled. Agents showed up in nearly half of active public Slack channels, adding to human conversation rather than replacing it. Staff with a year or more of tenure were 50% more likely each month to move departments in 2026 than in 2025.
That has a security and quality edge as well as an efficiency one. If no one can read everything the agents produce, some output ships unread. We have seen what that leads to: in our report on agents that claim success they did not achieve, the fix was always a verification step a human trusted. The more agents you run, the more that step has to be designed, sampled and automated rather than left to whoever has a spare minute.
5. The same bottleneck, in public: Postgres 19 slips
If you want to see the review bottleneck outside one AI lab, look at PostgreSQL, one of the most widely used databases in the world. The Register reported on 16 October that the PostgreSQL Global Development Group has pushed version 19 from its usual late-September or early-October slot to around 19 November, and pulled between eight and ten significant features, which are now expected in version 20.
Peter Eisentraut, a leading PostgreSQL committer and a vice president at EDB, explained in a 9 October blog post that LLM-assisted code review is part of the reason. Some pulled features had architectural problems that could not be fixed safely during the beta. But AI reviewers also surfaced a stream of what he called “relatively harmless issues involving various edge cases.” In one case, a feature drew around 40 separate defect reports, each needing a small fix, and the volunteers ran out of time. His point was that AI review got much better after the features were written, which effectively raised the quality bar mid-flight. He was clear that the reverts were not caused by AI-written code. Support for SQL property graph queries was among the casualties; a useful table-maintenance feature, REPACK CONCURRENTLY, survived.
This is the same pattern we described when Bank of America hired 1,000 contractors to fix what AI found, and when a flood of AI-found reports pushed Google to pause its open-source bug bounty. Generation and detection have become cheap. Deciding, fixing and signing off have not.
6. What this means for your business
Budget for review capacity, not just licences. If agents multiply output by five, the people who check that output need tools, time and clear rules for what gets sampled versus fully read. Otherwise the gains turn into a queue, or into unreviewed work reaching customers.
Rethink, rather than abandon, junior roles. Cutting entry-level hiring saves money this year and creates a senior shortage later. A better model is to hire juniors into review, testing and agent-supervision work, where they learn judgment by checking agent output against real outcomes.
Treat standing instructions as company assets. Anthropic’s 8,500 standing instructions are, in effect, a new kind of internal documentation and policy. They need owners, version control and security review, because an instruction file can grant an agent more reach than intended. Our guide to treating AI like an insider threat covers the controls.
Change what you hire for. The shift Anthropic saw, towards autonomy, judgment and customer focus, is a reasonable template. As we argued in you’re competing with people using AI, the advantage goes to people who can direct and check the tools, not just use them.
Give staff a share in the upside. The authors credit employee stock ownership with aligning staff with AI rollout, comparing it to the post-war bargain in which productivity gains were shared. Most firms cannot copy that exactly, but being explicit about how gains will be shared, and offering internal moves before cuts, reduces resistance.
7. The eight-question agent workforce audit
Ask these in your next leadership meeting. For each one, a good answer is described; if you cannot give it, that is your next project.
- How many agent hours do we run, and who directs them? Good answer: a rough monthly figure by team, with a named human owner for every agent or workflow.
- What share of agent output does a person actually read? Good answer: a stated sampling rate per workflow (for example 100% for customer-facing or financial changes, 10% for internal drafts) and a log showing it happens.
- Where is work waiting on us? Good answer: you can name the two or three approval steps where agent work queues longest, and someone is tasked with shortening them.
- Who owns our standing instructions? Good answer: prompts, rules and agent configuration files sit in version control, with review before changes and a check for secrets or over-broad permissions.
- Can we fix what our AI tools find? Good answer: a triage rule for AI-found bugs and issues (severity, owner, deadline), so a flood like Postgres’s does not stall a release.
- What happened to our entry-level pipeline? Good answer: junior hiring numbers for the last two years, and a plan for how new staff learn judgment if agents do the routine work.
- Have our job descriptions caught up? Good answer: hiring criteria that test for autonomy, judgment and the ability to verify AI output, not only task execution.
- Are we measuring quality as well as volume? Good answer: at least one quality metric (rework, reverts, defects, complaints) tracked alongside output, as Anthropic did with code deletions and reversions.
If you are giving agents their own accounts and identities, as Google now does with Gemini agents that get a Workspace account, questions 1 and 4 are also access-control questions: every agent identity needs an owner, a permission review and an off switch.
8. The bottom line
Anthropic’s own numbers show what an agent-heavy workplace looks like a year or two early: far more output per person, fewer junior hires, and people spending more of their day directing, reviewing and meeting. The lesson for every business is that buying AI capacity is now the easy part. The hard part is the human layer around it: who checks the work, who fixes what the tools find, and how the next generation of staff learns to do both. Firms that design for that will keep their gains. Firms that do not will find the agents waiting on them, about 80% of the time.
Frequently asked questions
What did Anthropic’s employment study find?
Published on 16 October 2026, it found that agents at Anthropic now work about six hours for every human hour in employee-directed sessions, code commits per engineer are about five times their level six months earlier, and the junior share of technical hires fell from 27% to 9% since early 2025.
Is AI replacing junior jobs at Anthropic?
Junior hiring has fallen sharply. Junior staff dropped from 21% to 14% of technical employees, and about a third of ML managers said they hire less because of AI. The authors say protections for current staff do not help people never hired, and call for retraining and income-support policy.
Did code quality drop when engineers used AI agents?
Not by the study’s measure. Comparing agent-written and human-written code from the same engineer in the same month, Anthropic found no difference in how often the code was later deleted or reverted. Adoption was voluntary, so the results are not from a randomised trial.
Why is human attention the bottleneck?
Interactive agent sessions at Anthropic are about 80% idle time waiting for a person to respond, and reading all agent output would take non-technical teams more than half a working week. Agents produce faster than people can review, decide and coordinate.
Why was PostgreSQL 19 delayed?
According to The Register, improved LLM-assisted code review surfaced many small edge-case defects late in the cycle, and some features had deeper design issues. Developers pulled eight to ten features into version 20, and release moved to around 19 November 2026.
What should businesses do in response?
Measure agent use and review rates, budget for human review and fixing capacity, put agent instructions under version control and security review, keep a path for junior staff through review and supervision work, and track quality metrics alongside output.