The landscape · Education
Education: what AI is really doing
Brief me in 2 minutes
AI in Education.
The two biggest US school systems just banned generative AI for most kids while quietly funding it for every teacher, which is the real split in this sector: adults keep the tools, children lose them pending a policy rewrite. Universities have simultaneously given up on catching AI cheating with detection software and are going back to blue books and oral exams instead. The strongest evidence that AI tutoring actually works comes from a narrow, high-motivation population (Harvard physics undergrads), not from the mass-market K-12 chatbots being marketed on the strength of it.
The game: we show you 6 things happening. You call each one real now, early or mostly hype. Then the receipts.
Shift 1 of 6
NYC and LA ban AI for students, not for teachers
Real now
Two of the three biggest US districts didn't announce a plan, they flipped the switch this month, for kids currently sitting in class.
How it works. New York City's Department of Education imposed a one-year moratorium on student-facing generative AI for grades 2-K through 8th, covering roughly two-thirds of its enrollment, while still allowing five limited pilots for up to 50,000 high schoolers and letting teachers keep using AI for lesson planning and translation. Los Angeles Unified went further days later, blocking all students district-wide on district-issued devices while a new Generative AI Ad Hoc Committee reviews policy.
Overstated: The 'ban' headlines overstate a blanket rejection of AI in schools: NYC explicitly carved out approved teacher tools and five high-school pilots, and LAUSD's restriction is framed as temporary pending committee recommendations, not a permanent policy.
Shift 2 of 6
Universities give up on AI detectors and bring back blue books
Real now
This is a policy already changed at named universities with the vendor already switched off, not a working paper about what schools might do someday.
How it works. A growing list of universities, including Yale, Vanderbilt, Johns Hopkins, Indiana, Northwestern, Georgetown and NYU, have banned faculty reliance on AI-detection software or disabled Turnitin's AI detector outright, because false-accusation rates were damaging trust faster than the tools were catching cheating. In its place, faculty are reviving proctored blue-book exams and oral defenses.
Overstated: Detection vendors still advertise near-perfect accuracy, but universities that actually deployed the tools found 'humanizing' rewriting services reliably defeat them, and Berkeley's own large-scale study found cheating varies wildly by discipline and rises sharply with daily AI use, not the clean signal detector marketing implies.
Shift 3 of 6
AI tutors are being redesigned from answer machines into Socratic guides
Early
The redesigned, proactive Khanmigo is a summer 2026 rollout to district partners and the flagship RCT is 194 Harvard undergrads in one course, so call it promising, not proven at scale.
How it works. OpenAI's Study Mode, Khan Academy's redesigned Khanmigo, and startups like Wild Zebra and Third Space Learning's Skye all now withhold the final answer and instead walk a student through reasoning step by step, prompting for prerequisite gaps and adjusting difficulty in real time rather than answering on request.
Overstated: The most-cited evidence that AI tutoring works, a Harvard trial showing more than double the learning gains of an active-learning classroom, was run on 194 Harvard physics undergrads, a highly structured subject with a high-motivation, high-preparation population; it is not proof the same gains show up in a mass-market K-12 chatbot.
Shift 4 of 6
Teacher-tools went from pilot to near-default classroom software, mostly for free
Real now
Eight-figure rounds already closed and millions of teachers already opened the extension; whatever the true weekly-use number is, the product is live in classrooms, not on a roadmap.
How it works. MagicSchool, Brisk Teaching and SchoolAI all crossed into mass adoption through free or freemium browser-extension distribution rather than district procurement cycles, handling lesson planning, differentiation, grading support and admin work, and each has raised a real venture round ($45M, $15M and $25M respectively) on the back of it.
Overstated: Signup and 'monthly use' counts are registration numbers, not measured classroom outcomes; Khan Academy's own disclosure that only 15% of students with Khanmigo access actually use it is a reminder that reach and adoption are not the same thing across this whole category.
Shift 5 of 6
Corporate L&D stops sending workers to 'training' and puts AI inside the job itself
Early
BCG and Bersin are describing where budgets are pointed and one workflow example each, not a roster of companies that have already retired their course library.
How it works. Instead of standalone courses, companies are embedding AI guidance directly into the tools employees already use, so instruction arrives at the moment it's needed rather than as a separate learning session; BCG frames this as L&D finally getting 'out of the classroom and into the workflow,' with corporate AI spending on track to roughly double in 2026.
Overstated: Consultancy-commissioned stats like 'companies at Level 4 are 10-times more likely to be innovation leaders' are correlational and self-reported maturity-model claims, not evidence that embedding AI in workflows causes better business outcomes.
Shift 6 of 6
AI-assisted cheating is now its own funded startup category
Real now
Cluely already has paying users and reported revenue, GPTZero already sits inside a million educators' grading workflow; both sides of this fight are shipping, not pitching.
How it works. Cluely packages a mainstream LLM with screen and audio capture to feed real-time hidden answers into interviews, exams and calls, has raised a $15M Series A from a16z on top of a $5.3M seed, and markets itself as a replacement for the moment someone would otherwise 'reach for ChatGPT.'
Overstated: Founder Roy Lee's line that 'the invisibility function is not a core feature' downplays exactly the capability, undetectable real-time answers during a screen share, that the product is built and marketed around.
The scary part
What happens to jobs in Education
What AI is already taking:
- Writing feedback and unit planning for teachers EdSurge ↗
- One-to-one math tutoring delivery, replacing a human tutor for the session Third Space Learning ↗
- Diagnosing a student's knowledge gaps and adjusting lesson difficulty in real time TIME ↗
- Delivering point-of-need job training inside the tools employees already use BCG ↗
What stays human:
Deciding whether a struggling kid needs more time or more challenge still sits with a teacher, which is exactly why Khan Academy rebuilt Khanmigo around teacher-set guardrails instead of letting the bot decide alone. And judging whether a piece of work is actually a student's own thinking still needs a human in the room, which is precisely why universities dropped the software that tried to automate that judgment and put people back in oral exams.
Get better · 15 minutes
Designing and grading an assessment that survives a student with a $20-a-month AI subscription is the scarce skill right now, not using AI yourself; it's what Yale, Vanderbilt and Northwestern are rewriting policy around because nobody has fully solved it yet.
Try this this week:
This week, take one assignment you'd normally hand a student and run it through ChatGPT's Study Mode (free at chatgpt.com) instead of plain ChatGPT, and count how many turns pass before it gives you the actual answer versus asks you a question back; that turn count is the exact thing districts are now grading tutoring vendors on.
Your move
Generic teacher lesson-plan generators and generic tutoring chatbots are crowded: MagicSchool, Brisk and SchoolAI each already claim over a million educators or classrooms. Open space is in the two problems this briefing shows are unsolved: assessment formats that survive an AI-equipped student at scale (the blue-book revival has no software layer yet), and workflow-embedded corporate learning that measures task performance instead of course completion.
Watch usage, not signups: Khan Academy's own disclosure that only 15% of students with Khanmigo access actually used it is the model for how registration and 'interaction' counts hide real adoption across this category. Red flag any pitch that cites the Harvard physics RCT's 'double the learning gains' as generalized proof, without disclosing it was 194 highly-motivated Harvard undergrads in a single structured subject.
If you run a school or district, test one unit through a Socratic-mode tutor (Khanmigo's redesign, Wild Zebra, or Skye) against your current approach this quarter and track completion and error rates, not logins. If you handle academic integrity, run your detection tool (Turnitin, GPTZero) against writing samples students have admitted were AI-assisted, to get a real false-positive/false-negative baseline before it's used in a misconduct hearing.
The skill in demand is designing and grading assessments that survive an AI-equipped student, oral defenses, in-person blue books, process logging, which is exactly what Yale, Vanderbilt and Northwestern are now rewriting policy around. First step: ask to sit in on your institution's academic-integrity policy revision and pull the specific rubric changes.
What's actually changing
- 1
NYC and LA ban AI for students, not for teachers
How it works. New York City's Department of Education imposed a one-year moratorium on student-facing generative AI for grades 2-K through 8th, covering roughly two-thirds of its enrollment, while still allowing five limited pilots for up to 50,000 high schoolers and letting teachers keep using AI for lesson planning and translation. Los Angeles Unified went further days later, blocking all students district-wide on district-issued devices while a new Generative AI Ad Hoc Committee reviews policy.
Why it matters. Two of the three largest US districts now treat AI as unsafe for children by default but useful for staff, which forces every K-12 edtech vendor selling student-facing tools to either exit the largest markets or rebuild as a teacher-only product.
What's overstatedThe 'ban' headlines overstate a blanket rejection of AI in schools: NYC explicitly carved out approved teacher tools and five high-school pilots, and LAUSD's restriction is framed as temporary pending committee recommendations, not a permanent policy.
NYC Department of EducationLAUSD
3 receipts
- NYC's moratorium covers roughly 600,000 students, about two-thirds of the system's enrollment, from grades 2-K through 8th.
nearly 600,000 public school students, or two-thirds of the system's total enrollment
NYC Mayor's Office, 2 Sept 2026 ↗ - NYC still permits a small high-school pilot population alongside the ban.
maximum of 50,000 high school students in general education classes, or approximately 5% of the public-school student body, five limited AI pilots
NYC Mayor's Office, 2 Sept 2026 ↗ - LAUSD's restriction is explicitly temporary while a committee reviews safeguards.
in place while the district reviews instructional use and safeguards.
K-12 Dive, 2 Sept 2026 ↗
- NYC's moratorium covers roughly 600,000 students, about two-thirds of the system's enrollment, from grades 2-K through 8th.
- 2
Universities give up on AI detectors and bring back blue books
How it works. A growing list of universities, including Yale, Vanderbilt, Johns Hopkins, Indiana, Northwestern, Georgetown and NYU, have banned faculty reliance on AI-detection software or disabled Turnitin's AI detector outright, because false-accusation rates were damaging trust faster than the tools were catching cheating. In its place, faculty are reviving proctored blue-book exams and oral defenses.
Why it matters. The industry's original answer to AI cheating, detection software, is being abandoned at the institutional level inside two years, pushing grading back toward labor-intensive formats that don't scale to large lecture courses.
What's overstatedDetection vendors still advertise near-perfect accuracy, but universities that actually deployed the tools found 'humanizing' rewriting services reliably defeat them, and Berkeley's own large-scale study found cheating varies wildly by discipline and rises sharply with daily AI use, not the clean signal detector marketing implies.
YaleVanderbiltJohns HopkinsIndiana University Kelley School of BusinessNorthwesternGeorgetownNew York UniversityTurnitinGPTZero
3 receipts
- Multiple top universities now ban or discourage faculty reliance on AI detectors.
A growing number of institutions—including Yale, Vanderbilt, Johns Hopkins and Indiana Universities—to enact academic integrity policies that ban or discourage faculty reliance on AI-detection programs.
Inside Higher Ed, 5 Aug 2026 ↗ - Faculty are shifting to in-person exam formats instead of software detection.
such as a blue book exam or an in-person oral exam
Inside Higher Ed, 5 Aug 2026 ↗ - A 95,000-student study found cheating with AI concentrated among frequent users and varies by discipline.
More than 95,000 students at 20 research-intensive public universities responded to questions about how they use AI, including whether they use it to cheat.
UC Berkeley News, 21 May 2026 ↗
- Multiple top universities now ban or discourage faculty reliance on AI detectors.
- 3
AI tutors are being redesigned from answer machines into Socratic guides
How it works. OpenAI's Study Mode, Khan Academy's redesigned Khanmigo, and startups like Wild Zebra and Third Space Learning's Skye all now withhold the final answer and instead walk a student through reasoning step by step, prompting for prerequisite gaps and adjusting difficulty in real time rather than answering on request.
Why it matters. Funders are now paying specifically for this behavior: Khan Academy rebuilt Khanmigo after admitting only a small share of students with access actually used it, and the Gates Foundation-backed Digital Promise program is funding an open-source tutor because existing ones 'give answers too quickly.'
What's overstatedThe most-cited evidence that AI tutoring works, a Harvard trial showing more than double the learning gains of an active-learning classroom, was run on 194 Harvard physics undergrads, a highly structured subject with a high-motivation, high-preparation population; it is not proof the same gains show up in a mass-market K-12 chatbot.
OpenAIKhan AcademyWild ZebraThird Space Learning
4 receipts
- Only 15% of students with Khanmigo access actually used it, prompting Khan Academy to redesign it to be proactive rather than chat-initiated.
Only 15 percent of students with access to Khanmigo actually use it
EdTech Innovation Hub, 5 May 2026 ↗ - Wild Zebra's tutor is built to draw out reasoning rather than hand over answers.
Rather than handing over the answer, the platform works students toward it with questions, drawing out the reasoning step by step.
GeekWire, 6 Aug 2026 ↗ - A Harvard RCT found AI-tutored students learned significantly more than a classroom taught with active learning, but on a narrow, high-achieving population.
Students using the AI tutor learned significantly more, with median learning gains more than double those of students in the traditional active learning classroom.
Impact AI Education (Substack), 1 Jan 2026 ↗ - The Gates Foundation-backed grant exists because current AI tutors give answers too fast and miss motivation signals.
often give answers too quickly, talk too much, miss signs of student motivation, and do not always create enough space for students to work through problems
EdTech Innovation Hub, 1 June 2026 ↗
- Only 15% of students with Khanmigo access actually used it, prompting Khan Academy to redesign it to be proactive rather than chat-initiated.
- 4
Teacher-tools went from pilot to near-default classroom software, mostly for free
How it works. MagicSchool, Brisk Teaching and SchoolAI all crossed into mass adoption through free or freemium browser-extension distribution rather than district procurement cycles, handling lesson planning, differentiation, grading support and admin work, and each has raised a real venture round ($45M, $15M and $25M respectively) on the back of it.
Why it matters. This, not student-facing chatbots, is where AI has actually reached scale in US classrooms: MagicSchool counts over 6 million educator signups across 10,000+ schools, Brisk over 1 million educators in 100 countries, and SchoolAI over 1 million classrooms in 400+ districts.
What's overstatedSignup and 'monthly use' counts are registration numbers, not measured classroom outcomes; Khan Academy's own disclosure that only 15% of students with Khanmigo access actually use it is a reminder that reach and adoption are not the same thing across this whole category.
MagicSchoolBrisk TeachingSchoolAI
3 receipts
- MagicSchool raised a $45M Series B and counts over 6 million signed-up educators across 10,000+ schools.
$45 million funding round, led by Valor Equity Partners
MagicSchool, 11 Feb 2025 ↗ - Brisk Teaching reports 1 million educators across 100 countries and 2,000+ schools, saving teachers up to 10 hours a week, after a $15M Series A.
In just over a year, Brisk has grown to serve over one million educators across 100 countries and partnered with more than 2,000 schools and districts.
Bessemer Venture Partners, 25 Mar 2025 ↗ - SchoolAI is used in over 1 million classrooms across all 50 states and 80+ countries, backed by a $25M Series A from Insight Partners.
SchoolAI is now used in more than 1 million classrooms across all 50 U.S. states and over 80 countries. More than 400 school districts have adopted the platform
Just AI News, 1 June 2025 ↗
- MagicSchool raised a $45M Series B and counts over 6 million signed-up educators across 10,000+ schools.
- 5
Corporate L&D stops sending workers to 'training' and puts AI inside the job itself
How it works. Instead of standalone courses, companies are embedding AI guidance directly into the tools employees already use, so instruction arrives at the moment it's needed rather than as a separate learning session; BCG frames this as L&D finally getting 'out of the classroom and into the workflow,' with corporate AI spending on track to roughly double in 2026.
Why it matters. This reroutes L&D budgets from course libraries and completion metrics toward workflow-embedded tools measured by task performance, changing who edtech-adjacent startups need to sell to inside a company.
What's overstatedConsultancy-commissioned stats like 'companies at Level 4 are 10-times more likely to be innovation leaders' are correlational and self-reported maturity-model claims, not evidence that embedding AI in workflows causes better business outcomes.
BCGThe Josh Bersin Company
3 receipts
- BCG describes AI moving corporate learning out of the classroom and into the flow of work.
For years, L&D has been trying to get out of the classroom and into the workflow. AI has done that for them.
BCG, 23 June 2026 ↗ - Companies plan to roughly double AI spending in 2026 to about 1.7% of revenue.
companies plan to double their spending on the technology in 2026, accounting for about 1.7% of revenues
BCG, 23 June 2026 ↗ - Josh Bersin's research puts the corporate training market at $400 billion and cites most companies as unable to keep up with new-skill demand.
74% of companies tell us they are not keeping up with their company's demand for new skills
Josh Bersin, 11 Feb 2026 ↗
- BCG describes AI moving corporate learning out of the classroom and into the flow of work.
- 6
AI-assisted cheating is now its own funded startup category
How it works. Cluely packages a mainstream LLM with screen and audio capture to feed real-time hidden answers into interviews, exams and calls, has raised a $15M Series A from a16z on top of a $5.3M seed, and markets itself as a replacement for the moment someone would otherwise 'reach for ChatGPT.'
Why it matters. The same underlying model capability now funds three adversarial layers at once, cheat-assist tools, detectors, and 'humanizer' rewriters that defeat detectors, turning academic integrity into a live arms race with venture money on every side.
What's overstatedFounder Roy Lee's line that 'the invisibility function is not a core feature' downplays exactly the capability, undetectable real-time answers during a screen share, that the product is built and marketed around.
CluelyGPTZero
4 receipts
- Cluely raised a $15M Series A from Andreessen Horowitz to build a hidden-window AI assistant for interviews and exams.
$15 million Series A
TechCrunch, 9 July 2025 ↗ - That Series A followed an earlier $5.3 million seed round that had already drawn backlash for normalizing cheating.
Cluely, an AI startup helping users cheat during exams, interviews, and even sales calls, has raised $5.3 million in seed funding despite sparking backlash for normalizing deception.
eWEEK, 28 Apr 2025 ↗ - Cluely's founder frames the product as a replacement for reaching for ChatGPT in the moment.
Every time you would reach for chatgpt.com, our goal is to create a world where you instead reach for Cluely.
TechCrunch, 9 July 2025 ↗ - GPTZero markets near-perfect detection accuracy to over a million educators.
99% Accuracy 17M+ Users 1M+ Educators
GPTZero, 23 Sept 2026 ↗
- Cluely raised a $15M Series A from Andreessen Horowitz to build a hidden-window AI assistant for interviews and exams.
Who's doing it
11 companies to know, 8 with a round we can source. Newest first where dated.
Search these companies in the startup database ↗ · Explore all research tools
Third Space LearningSpoken AI math tutoring (UK)
UK-based tutoring provider whose spoken AI tutor Skye talks and listens to pupils through math problems in real time, modeled on tutor pedagogy.
Latest · 23 Sept 2026Skye has now taught over 196,000 students across 4,300+ UK schools. Third Space Learning ↗
GPTZeroAI-writing detection
AI content detector used by schools, hiring and publishing to flag text believed to be generated by large language models.
Latest · 23 Sept 2026Used in over 3500 colleges, claiming a 99% accuracy rate against human writing. GPTZero ↗
Wild ZebraK-12 AI tutoring (math and reading, grades 2-9)
AI learning platform for grades 2-9 math and reading that guides students to answers through questions instead of giving them directly.
Latest · 6 Aug 2026Raised an oversubscribed $6M seed round led by Trilogy Equity Partners, now reaching tens of thousands of students. GeekWire ↗
$6M Seed, 6 Aug 2026 source ↗
ProLearnK-12 and exam-prep AI tutoring (India)
Real-time, conversational AI learning companion for Indian K-12 students and competitive-exam aspirants (JEE, NEET, UPSC, CAT).
Latest · 2 June 2026Raised a pre-seed round to build a real-time interactive AI learning companion ahead of public launch. Entrepreneur India ↗
INR 30 crore Pre-seed, 2 June 2026 source ↗
StudyFetchStudent study tools (flashcards, quizzes, AI tutor)
AI-powered study platform generating flashcards, quizzes, summaries and an AI tutor grounded in students' own course material.
Latest · 5 Jan 2026More than 6 million students used StudyFetch in 2025, with weekly active usage up over 50% week-over-week during December finals. PR Newswire ↗
$11.5M Series A, 1 June 2025 source ↗
Squirrel AIAdaptive K-12 tutoring, US expansion
Adaptive AI tutoring system that diagnoses knowledge gaps and customizes lessons in real time; the dominant player in China now entering the US.
Latest · 1 Jan 2026Entering the US market in 2026 to test whether its adaptive system adapts to American curricula, after reaching 52 million students in Asia. TIME ↗
CluelyAI cheating / real-time assistance
Real-time AI assistant that reads a user's screen and audio through a hidden window to feed answers during interviews, exams and calls.
Latest · 9 July 2025Raised a $15M Series A led by Andreessen Horowitz on top of an earlier $5.3M seed. TechCrunch ↗
$15M Series A, on top of an earlier $5.3M seed Series A, 9 July 2025 source ↗
SchoolAIClassroom AI platform (teachers and students)
Classroom Experience platform giving teachers AI assistants for prep, lesson planning and admin work alongside student-facing tools.
Latest · 1 June 2025Now used in over 1 million classrooms across all 50 US states and 80+ countries, with 400+ districts adopting the platform, after a $25M Series A. Just AI News ↗
$25M Series A, 1 June 2025 source ↗
Brisk TeachingTeacher productivity tools
Browser-extension AI assistant that automates grading, tutoring support and resource creation directly inside teachers' existing tools.
Latest · 25 Mar 2025Reached over one million educators across 100 countries and 2,000+ schools, saving teachers up to 10 hours a week, after a $15M Series A. Bessemer Venture Partners ↗
$15M Series A, 25 Mar 2025 source ↗
MagicSchoolTeacher productivity tools
AI co-pilot for teachers covering lesson planning, differentiation and adaptive one-on-one tutoring support for students.
Latest · 11 Feb 2025Raised a $45M Series B and passed 6 million signed-up educators across 10,000+ schools in 160 countries. MagicSchool ↗
$45M Series B, 11 Feb 2025 source ↗
Your move
Our read, built on the receipts above.
Generic teacher lesson-plan generators and generic tutoring chatbots are crowded: MagicSchool, Brisk and SchoolAI each already claim over a million educators or classrooms. Open space is in the two problems this briefing shows are unsolved: assessment formats that survive an AI-equipped student at scale (the blue-book revival has no software layer yet), and workflow-embedded corporate learning that measures task performance instead of course completion.
Watch usage, not signups: Khan Academy's own disclosure that only 15% of students with Khanmigo access actually used it is the model for how registration and 'interaction' counts hide real adoption across this category. Red flag any pitch that cites the Harvard physics RCT's 'double the learning gains' as generalized proof, without disclosing it was 194 highly-motivated Harvard undergrads in a single structured subject.
If you run a school or district, test one unit through a Socratic-mode tutor (Khanmigo's redesign, Wild Zebra, or Skye) against your current approach this quarter and track completion and error rates, not logins. If you handle academic integrity, run your detection tool (Turnitin, GPTZero) against writing samples students have admitted were AI-assisted, to get a real false-positive/false-negative baseline before it's used in a misconduct hearing.
The skill in demand is designing and grading assessments that survive an AI-equipped student, oral defenses, in-person blue books, process logging, which is exactly what Yale, Vanderbilt and Northwestern are now rewriting policy around. First step: ask to sit in on your institution's academic-integrity policy revision and pull the specific rubric changes.
What's next
- Nov 2026 Digital Promise's Gates Foundation-backed $8M open-source K-12 math tutor grant begins work edtechinnovationhub.com ↗
- June 2027 LAUSD's Generative AI Ad Hoc Committee delivers policy recommendations to the school board k12dive.com ↗
Other industries
- FinanceBanks move agentic AI from chatbots to running back-office and dealmaking work
- Software engineeringSpaceX buys the leading AI coding startup outright, for $60 billion
- HealthcareAmbient scribes go from pilot project to clinical infrastructure
- LegalLegal AI vendors go from research assistant to agent that produces the deliverable
- Real estateLeasing and resident calls move to voice AI agents that never sleep
- Sales and marketingCustomer support agents now get paid per resolved ticket, not per seat
- Jobs, HR and recruitingAI runs the first-round interview, not just the resume filter
- InsuranceClaims move from adjuster-first to AI-first, humans supervise the exceptions
- CybersecurityAutonomous AI agents take over penetration testing and vulnerability hunting
- Consulting and accountingAI-written client reports become a recurring, public Big Four embarrassment
- Retail and e-commerceShopping agents get a checkout button, and Amazon spends 2026 deciding who's allowed to press it
- Media and creativeAI video splits into model labs that train and aggregators that resell
- Manufacturing and roboticsHumanoid robots move from demo livestreams to paid warehouse shifts
- Transport and robotaxisRobotaxis go from demo to metered fare across a dozen-plus US cities
- Energy and data-centre powerGrid rules get rewritten mid-buildout, state by state
- GovernmentGSA turns AI adoption into a government-wide subscription, then flips 'free' to metered
- Agriculture and foodAutonomous tractor fleets learn to run the tools behind them, not just the wheel