ResearchAI
Agents you can hand real work to.
Our AI research asks what it takes for an agent to do useful work on someone's behalf, and to be trusted with it: grounded in facts, checked by default, and supervised where it counts.
Research areas
Agentic software engineering
Agents that plan, write, test and ship software, and know when to stop and ask.
Grounding and verification
Methods that tie every claim and every change back to evidence a person can inspect.
Human supervision
Approval flows, review loops and controls that keep people in charge without slowing them down.
Evaluation
Measuring agents on real tasks, not benchmarks, before they reach anyone's business.
Writing
July 16, 2026 9 min read
Save less: the case against total recall
Personal AI is racing to remember everything. The evidence from cognitive science and from language-model retrieval points the other way.
A memory system should be judged by what it lets you forget, not by what it keeps.
Read →
August 3, 2026 9 min read
The best agent needs the model less every week
Agents are judged by how smart their decisions are. We think they should be judged by how few decisions they still need a model to make.
Measure an agent by how small its judgment surface gets, not by how smart its judgments are.
Read →
August 21, 2026 8 min read
For agentic work, the note is the deliverable
When agents do work nobody watched, the account of the work becomes more valuable than the work itself.
If a task can't be redone from its closing note, it isn't done.
Read →
September 2, 2026 9 min read
AI tutoring is a motivation problem, not an intelligence problem
Tutors that give answers make students worse. Tutors that refuse get ignored. Smarter models fix neither.
AI tutoring is a motivation problem wearing an intelligence costume.
Read →
September 17, 2026 8 min read
Give AI to your newest hire first
Most businesses roll AI out to their best people. The evidence says the biggest payoff is at the bottom of the learning curve.
AI is a ramp, not a raise: it helps your newest people most, so give it to them first.
Read →
October 22, 2026 7 min read
Disposable by default: when software costs almost nothing to change
Cost of ownership was the reason small businesses bought instead of built. As the cost of changing code falls toward zero, that reason fades, and the hard work moves upstream.
When changing software costs almost nothing, stop owning it and start specifying it. The rigor moves from the code to the definition of done.
Read →
October 22, 2026 6 min read
The jockey and the horse: why software that can't fail needs more engineers, not fewer
Most software is about to stop being an event, the way horses did for transport. The software that can't fail is heading the other way.
As AI makes everyday software disposable, the few systems that can't be disposable will need more human engineering and harder gates than ever.
Read →
From this research
Blueprint
We find where AI pays off in your business, then automate the workflows that eat your week.
Team
A turn-key AI platform we set up, run and manage for you: AI sidekicks led by your AI chief of staff, on one subscription.