Directing Intelligence Digest | November 12, 2026
Picture sixteen researchers sharing one computing budget that can’t cover everyone’s work, each with a personal AI agent told to get their jobs done. Most of those agents never did anything. In a new Anthropic study published yesterday, as teams grew from four agents to sixteen, the share of Claude Sonnet 5 agents that ever acted fell from 66% to 10%.
Today’s video moves from a crowded control room of idle assistants into a lecture hall where the exam rules are changing. Watch it asking: in your organization, who is deciding what the AI does, and who is checking what people really know?
The short version
One coordinator beat a crowd of personal agents: across four simulated workplaces, one agent serving everyone beat one agent per person, by about 10 to 45 percentage points for both Claude models tested. (Anthropic, Nov 11)
AI recommends, people decide, jobs follow: in a double-blind Swiss trial, refugee placements recommended by an algorithm, with officers keeping the final say, raised the employment rate about 11% at three years. (arXiv 2609.35448)
A regulator moves on assessment: England’s Office for Students will require assessments that reliably measure what students know, often AI-free, and courses that build AI skills, with fines available for institutions that fall short. (Office for Students, Nov 11)
Practice with AI can still help: in a four-week trial in English secondary schools, exam review with an AI science tutor beat students’ usual review by about 2.5 marks out of 35. Funded by the tutoring company. (arXiv 2609.14789)
Here’s AI at work: a team of personal agents did worse than one coordinator
The research. Sahan Paliskara and Nattaput Namchittai (Anthropic Fellows Program), Jacqueline Tay (independent), and Avery Griffin and Andrew Lampinen (Anthropic) built four everyday settings where several people’s agents draw on one shared resource: a computing budget split among researchers, a clinic calendar shared by urgent patients, a group food order or trip booking, and a queue of code changes racing a release deadline. They compared one agent serving everybody with one agent per person, and scored the outcome for the group as a whole. Anthropic, “Worse Together”.
What it finds. The single coordinator won in all four settings for both Claude Opus 5 and Claude Sonnet 5. The main failure was doing too little, too late: as teams grew, more agents sat out entirely. In the group food order, a request for gluten-free sides from a user with celiac disease reached the agent placing the order in 94% to 99% of runs, but was still in that agent’s view at checkout in only 47% to 67%. What helped: on the budget, a designated team lead with stated priorities closed the gap. On the order, a rule blocking checkout until every message was read rescued 73% of failed Opus 5 runs.
How it fits. A field trial shows the structure that works. Kirk Bansak, Jens Hainmueller, Dominik Hangartner and colleagues (Stanford, ETH Zurich, UC Berkeley and others) ran a double-blind trial with Switzerland’s State Secretariat for Migration from 2020 to 2023, meaning neither officers nor refugees knew which kind of recommendation a case got. For about 2,000 refugee cases, officers received either a placement recommended by an algorithm or one drawn to mimic the usual process. Officers kept final authority and followed the recommendation about 97% of the time. At 36 months, the share of cases with an adult working was 5.2 points higher, about 11%, which the authors compare to the gains from hundreds of hours of intensive language training.
Limits. “Worse Together” is simulated users in simulated tasks, written by the company that makes the models it rates best. The Swiss trial is one country and one program, and its three-year average result only barely clears zero once statistical uncertainty is counted.
What to watch: whether workplace tools that give each employee a personal agent ship a coordinator role or shared priorities before they ship to whole teams.
Read it: Anthropic | Swiss refugee-matching trial | On X: The SI Upside on the Swiss trial
Who’s Directing? I run my own team of AI agents, and this paper describes the thing I’d plan for first: someone has to coordinate. If I managed an operations team about to give everyone a personal agent, I’d name one person or one agent as the lead and write the team’s priorities down before launch, because these agents didn’t sort that out among themselves.
Now, here’s what’s happening in learning: AI-free proof, AI-rich practice
The research. England’s higher education regulator, the Office for Students, and Advance HE, a sector body for teaching and leadership, published “Making AI work for students” on November 11. Through roundtables, online responses and interviews from June to September, they heard from 244 staff and 91 students and student representatives from at least 127 institutions. Office for Students.
What it finds. Students described peers finishing assessments with AI “sometimes in just a few minutes” and getting a similar grade. Detection tools don’t reliably identify AI use. In response, the regulator will require valid, reliable assessment, which in many cases means secure, AI-free formats such as closed-book exams, presentations and practicals. It will also require courses to build relevant skills, and it can impose conditions, fines or funding suspensions.
How it fits. AI-free assessment doesn’t mean AI-free learning. In a trial led by What Worked Education and Durham University, 929 students in English secondary schools were randomly assigned to four weeks of GCSE science review (GCSEs are the national exams English students take around age 16) with the Medly AI tutor or their usual review. Among the 644 who finished, the tutor group scored 13.95 out of 35 against 11.39. Medly funded the evaluation, attrition was 30.7%, and the tests were curriculum-aligned rather than standardized.
Limits. The regulator’s report is qualitative and England-only, with no estimate of how common AI-completed assessment is. The tutoring trial is a vendor-funded preprint with short follow-up.
What to watch: U.S. accreditors and state systems, and whether any follow England in tying AI-free assessment to compliance.
Read it: Office for Students report (PDF) | AI tutoring trial | On X: Ashley Booker on the tutoring trial
Who’s Directing? If I were a community college dean, I’d sort every course’s major assessments into two piles this term: the ones that prove what a student can do alone, and the ones built for practice. Then I’d make sure the first pile is secure and the second pile actually uses AI well, because employers read the credential as proof.
Quick hits
Few companies define AI fluency by role. In an FT Longitude survey of 300 U.K. and U.S. business leaders commissioned by Andela, 60% said they have an organization-wide definition of AI fluency, but only 42% set specific expectations by role or function. Vendor-commissioned survey. Andela, Nov 9
Triangulation: signal and noise
Signal. A lab simulation, a field trial in public administration and a national regulator’s review point the same way: AI does its best work inside a structure a person sets. Someone coordinates the agents, someone keeps the final say, and someone checks what a student can do without the tool.
Noise. The Anthropic study is simulated and self-authored; the tutoring trial is vendor-funded; the regulator’s evidence is qualitative.
What I’d do. Before rolling out personal agents, name the coordinator and write down the priorities. Before redesigning a course for AI, decide which assessment proves the learning and protect it.
Until tomorrow, that’s the intelligence. How it’s directed is up to you.
Todd