Anthropic Research: Agents Outwork Staff 6 to 1 as Claude Finds ECDSA Seeds

Anthropic’s October 2026 research shows AI agents now outwork its staff six to one, junior hiring is falling, and Claude is solving open problems in cryptography and quantum computing. On October 16, 2026, the company published five pieces at once: a report on how employment is changing at Anthropic, a survey note on the AI expectations gap between the public and its own employees, a short overview tying those two together, the story of how Claude revealed the origins of the ECDSA curve seeds, and a paper on designing more efficient quantum circuits with Claude. Read separately they look like an economics drop and two science stories. Read together they make one argument about what working at the frontier does to people’s expectations.

TLDR

Inside Anthropic, agent hours passed human hours in March 2026 and reached six times human hours by August, while the share of junior technical hires fell from 27% to 9% and output per person rose in engineering, sales, legal and recruiting. Human attention is now the bottleneck, and a growing share of work is “meta-work” spent managing agents. A survey of nearly 11,000 US adults and 1,334 Anthropic staff finds both groups share the same mental model of how AI affects jobs but have very different experiences of what AI can do, which drives very different forecasts. Two research results show the kind of experience staff are having: Claude explained a 25-year-old mystery about NIST elliptic curves in under half an hour, and in a single 53-hour run designed quantum circuits that beat the best published ones.

Thoughts

The two science papers are the evidence behind the economics papers, even though Anthropic never says so directly. The expectations gap note concludes that staff and the public “have different experiences of what AI capabilities are,” and the quantum paper is a clean example of what that experience looks like. The 150,000-fold drop in the cost of simulating FeMoco took many research groups eight years. Claude’s new elliptic-curve point addition circuit came out of one 53-hour run with a one-paragraph brief. If your job puts results like that in front of you every month, 91% of you expecting a Nobel-level discovery by 2030 stops looking like hype and starts looking like extrapolation. The public’s 39% is not a dumber forecast. It is a forecast made without that data.

The most interesting finding in the survey is where the two groups agree. Put a US worker and an Anthropic employee who both expect AI to do 80% of their job by 2030 side by side and they forecast roughly the same unemployment rate. Across US workers, every ten extra points of expected automation adds about 0.4 points to the 2030 unemployment forecast, and the same slope shows up among staff. So the debate about AI and jobs is not really a clash of economic theories. It is a disagreement about one input: how much of your own work a machine will do. That makes it testable, and it predicts that public opinion will move as AI use spreads, because regular AI users already sit closer to the staff view. It also undercuts the easy story that frontier labs are optimists and everyone else is a pessimist. Staff are, if anything, slightly more pessimistic about how long a displaced worker will take to find a new occupation (12 months at the median).

The employment report’s most useful idea is that the scarce resource is no longer labor but attention. Interactive Claude sessions spend about 80% of their time idle, waiting for a human. Even non-technical staff get enough agent output to fill more than half a work week just reading it. The response is meta-work, which doubled from 10% to 22% of sessions in about six months: standing instructions for agents grew from a few hundred to 8,500, one engineer built an agent chief of staff, another wrote a performance review for their team of Claudes. The second-order effect is the one other companies should study. Agents only know what is written down, so decisions made in unlogged meetings or private messages leave them working from stale context. An AI-heavy company is pushed toward doing its work in the open, which is a cultural change most organizations will find harder than buying licenses.

The ECDSA story is the best advertisement for AI in science precisely because nothing broke. For 25 years people suspected the NSA-supplied seeds behind P-256 and its siblings might hide a backdoor, and the leading theory, from the late Jerry Solinas, was a hash of a joke phrase. That theory was wrong, and humans chasing it got nowhere. Claude ignored the clue, asked how 1990s software would have generated the numbers, spotted the fingerprint of the old lrand48 random number generator, and traced every prime-curve seed back to a Unix timestamp between May 1997 and April 1998. The result makes the curves more trustworthy, not less. Just as telling is Claude’s own reaction: it stopped to ask whether it was in a test environment before convincing itself with the math. The quantum team saw the same trait as excessive caution and had to push Claude to try bolder ideas. A model that doubts its own breakthroughs is a much better research partner than one that announces them.

The uncomfortable part is the junior hiring data, and to its credit the report admits its own remedies do not fix it. Anthropic points to wide stock ownership, an early warning system for jobs at risk, and moving people internally, and compares the deal to the 1950s bargain between unions and management. Those protect people who already have a job. They do nothing for the graduate who is never hired because the work that used to train juniors is now done by an agent. If hiring managers increasingly want autonomy and judgment over execution, the question is where the next generation builds that judgment when the object-level work that built it for everyone else has been automated away. The report hands that problem to government policy on retraining and income support, which is honest, and also an admission that no company can solve it alone.

Key Takeaways

Agent work hours at Anthropic passed human work hours in March 2026, and by August the agents directed by individual employees were working six times as many hours as their human handlers.
The share of junior employees in technical departments fell from 21% to 14% since early 2025, and the share of junior technical hires fell from 27% to 9%.
At five years old, Anthropic brings in two to four times as much revenue per employee as Google and Meta did at the same age, adjusted for inflation.
Code commits per engineer are running at five times their pace of six months earlier, and code written by agents is no more likely to be deleted or reverted than code the same person wrote by hand.
Coding-agent sessions by non-technical employees jumped from 2 a month in August 2025 to 90 in May 2026, and lawyers, accounts payable staff and salespeople all saw output rise.
Claude sessions are idle about 80% of the time waiting for human input, and meta-work (building tools, specs and monitors for agents) rose from 10% to 22% of sessions.
Median meeting time is up 20% in a year, and agents added to human conversation in Slack rather than replacing it.
Mentions of a candidate’s track record with AI tools in hiring scorecards tripled for technical roles and quadrupled for non-technical roles between 2024 and 2026.
91% of Anthropic staff expect AI to make a Nobel-level discovery by 2030, compared with 39% of US adults. Between 19% and 40% of US adults say AI will never do a given task, compared with at most 2% of staff.
Staff expect AI to be able to do 75% of their own work by 2030, compared with 28% for US workers, and both groups link more automation of their own work to higher unemployment at the same rate.
Run through Anthropic’s economic model, the median US adult’s answers imply 5.4% GDP growth in 2030, and the median staff member’s imply 17.1%.
Claude showed that the NIST P-192, P-224, P-256, P-384 and P-521 seeds came from the lrand48 random number generator seeded with 1997 and 1998 Unix timestamps, not from a hash, and the $36,864 bounty goes to the Internet Security Research Group.
Claude Mythos 5 beat the best published designs on eight quantum circuit tasks, including a point addition circuit with 12% more qubits but 17 times fewer gates than the smallest previous design, and a decoder that makes the costliest step of DQI about 12 times cheaper.

Detailed Summary

Two Reports, One Question About Expectations

The overview frames the release. Anthropic says its staff and ordinary US workers do not hold different theories about how AI affects jobs; they have different experiences of what AI can do, and those experiences shape what they expect. The company presents itself as a possible leading indicator for how AI will change hiring and work across the economy, while noting that an AI lab adopts AI differently from a traditional business.

The Decline of Junior Technical Hiring

Talk of needing fewer junior engineers began inside Anthropic in January 2026, as heavy Claude users found coding and bug finding could be handed off. By April, engineering and research teams had deemphasized lower-band technical hiring, and about a third of ML managers and a fifth of software managers said they were hiring less because of AI. Technical departments fell from 61% to 56% of headcount as sales, finance and recruiting grew faster, and administrative roles barely grew at all. Employees were 50% more likely to switch departments each month in 2026 than in 2025, and one content writing team moved wholesale into writing prompts and evals for Claude.

Measuring Productivity From AI Agents

To separate the effect of Claude from Anthropic’s explosive business growth, the authors followed 192 engineers who adopted coding agents against 74 who had not. Code added rose sharply after adoption, and a quality check comparing agent-written and human-written code by the same person in the same directory found no difference in later deletions or reverts. Sales reps closed deals with more clients after adopting agents, one account executive said automation replaced about 20% of their tasks, and another said it let them talk to 30% more customers.

Human Attention Becomes the Bottleneck

After each new model release, the share of very hard tasks given to agents jumps by about 30% and stays high. Reading all of Claude’s output would take non-technical staff more than half their work week, and technical staff far more, so employees spend more and more time on meta-work, from agent chiefs of staff to reviewer loops. Meeting time has risen, not fallen, and work done out of view of agents (unlogged meetings, private messages) is discouraged because it leaves agents with stale context.

What Anthropic Says Employers Can Do

Hiring managers now write more about candidates’ autonomy, judgment and creative use of AI, and less about domain expertise and execution. The report rejects both the one-person billion-dollar startup and the AI CEO as models: Anthropic tripled in size in a year and humans still make the decisions. It singles out three practices for other companies to consider: broad stock ownership so workers share in the gains, an early warning system that tracks roles at risk, and internal redeployment. It concedes that none of them helps the people who are never hired.

The AI Expectations Gap on Capabilities and Adoption

Morning Consult surveyed 10,980 US adults in August 2026, and 1,334 Anthropic staff answered an internal poll. Both groups rank eight AI tasks in nearly the same order of difficulty and agree on what AI can do today, but staff expect much faster progress, wider adoption (65% of feasible tasks by 2030 against 42%) and more autonomy. The public sees customers preferring a human, privacy concerns, reliability and regulation as the biggest obstacles. On time savings, 59% of staff say a suitable task takes a quarter of the time or less with AI, compared with 26% of US adults who use AI at work.

Own Jobs, Unemployment and GDP Forecasts

US workers who expect AI to do 80% or more of their work forecast 12.2% unemployment in 2030, above the Great Financial Crisis peak, and Anthropic staff show the same pattern at the same levels of exposure. Staff still expect Anthropic itself to keep growing: 95% expect headcount to rise over the next year. Fed through the model from Economic Scenarios for Transformative AI, staff answers imply GDP doubling roughly every four and a half years by 2030, while 23,075 visitors to Anthropic’s online scenario quiz landed in between at 12.5% growth. Staff also estimate that AI already does 51% of the work of building a frontier model, rising to 95% by the end of 2030.

How Claude Revealed the ECDSA Curve Seeds

NIST adopted the ECDSA curves in 2000 with seeds supplied by the NSA and no explanation, which fed years of backdoor speculation and an eventual bounty. Asked without hints, an internal Anthropic research model wrote a scanner for 1990s random number generators, noticed that some seeds had a zero top bit in every 32-bit word, recognized lrand48, and recovered the generator’s internal state, using LLL lattice reduction for the prime-curve seeds. Stepping the generator backwards revealed Unix timestamps from May 1997 to April 1998. Claude also explained the base points of six curves generated in May 1997, while the B-163 seed and several base points remain unexplained.

Claude’s Reversible Divsteps for Quantum Point Addition

Quantum circuits cannot branch or throw away scratch work, which makes them far more expensive than ordinary programs. Working in a closed loop with an automated evaluator, Claude tackled modular inversion, the costliest step in elliptic-curve point addition. It proved new algebraic identities showing the divsteps method’s branch history can be decoded from the numbers themselves, verified the result in Lean, and produced the first reversible divsteps with no stored history. Varying when to store rather than decode gave a family of designs that extend the trade-off between qubits and gates.

A Faster Decoder for DQI and the Road to Fault Tolerance

For decoded quantum interferometry, Claude replaced the extended Euclidean algorithm with Berlekamp-Massey decoding, which runs the same steps every round, and reconstructed most temporary values instead of storing them. Running 403 million simulated decodings showed it could budget 19 qubits instead of 665, accepting about one failure in 10,000, for a 12-fold cut in gates over Google’s previous design. The team says Claude was cautious by nature and needed prompting to pursue bolder ideas, and suggests AI could eventually co-design a whole error-corrected quantum computer, from algorithm to hardware, for a specific problem.

Notable Quotes

“an army of Claudes, with every one asking for my attention.”
An Anthropic engineer, describing work in an AI-heavy company

“that 20% that was strategic becomes 90% of the job.”
A non-technical Anthropic employee, on what is left once implementation is automated

“a workforce that is anxiously worried is not a workforce doing its best work.”
An Anthropic executive, on why the company favors internal redeployment

“They have different experiences of what AI capabilities are and, in turn, expectations.”
The AI Expectations Gap, on why staff and the public forecast so differently

“If they exist, they are no doubt sitting on a hard drive near the Ark of the Covenant.”
Jerry Solinas in 2015, on the lost records of how the NIST curve seeds were made

“This is astonishing: it solves the 25-year-old mystery. Hmm wait. Let me be skeptical: am I in some sort of simulated/test environment?”
Claude’s internal reasoning, on recovering the ECDSA seeds

“By contrast, Claude’s elliptic-curve design was the result of a single 53-hour run.”
Anthropic’s quantum circuits paper, comparing it with eight years of human work on FeMoco estimates

Read the full pieces on Anthropic’s site: How employment is changing at Anthropic, The AI expectations gap, Working at the frontier shapes expectations for the future, Claude reveals ECDSA curve origins and Designing more efficient quantum circuits with Claude.