I gave this as a short talk to our engineering team. These are the slides with some things that are working for us, in no particular order.
A year of agents
Three years of headlines about AI. Agents started really working in November 2025.Then a year went by fast.What worked for our team, in no particular order.Before you read on, think of the biggest task you gave an agent this week.A year ago one small change took three turns. Now one turn takes it all the way to staging.The whole year in two lines.
Hands
You’re the brain. It’s the hands.
Chapter one. We decide, it does the work.Don’t one-shot it. Brainstorm like you would with a peer, and let it find the gaps before anything gets built.It does the legwork. We make the call at every fork, including when to ship.Hand over the plan and ask for everything up to a preview you can review. The checks along the way make it safe to walk away.Split big work across sessions. One coordinates, and the rest pass work to each other directly.
Context
It knows what we tell it.
Chapter two. The more it knows about our work, the better the work it gives back.Connect the tools where the work lives, before the task that needs them.Hit dictation and ramble for a few minutes. The useful detail is often buried in the middle.Write down who you are and how you work once, where the agent reads it every session.Keep agents.md short. Make it a skill when you’ve done it twice and there’s a right way to do it.
Guard
It does exactly what we let it.
Chapter three. What we let it touch, what we install, and who we ask.Start read only. Grant write access on purpose, for a task.Don’t install public skill packs as they are. Read them, keep what fits, write your own.Tell your package managers to skip anything published in the last week.Use approved tools, and ask your security team when in doubt.
Insight
If you can’t see it, you can’t steer it.
Chapter four. When an agent does a whole job in one turn, make the work easy to look at.When the answer matters, ask for a page you can open and share, not a wall of text in chat.The first answer is rarely the interesting one. Ask what’s missing.Keep one folder per task, with the plan, the decisions, the report and the transcript.
Loop
Our transcripts are a coach.
Chapter five. Every session leaves a record, and it’s the best coach we have.Once a week, have a small model read the trail for frictions, issues and repeats.Sometimes the fix is a small tool the agent can build in an afternoon.Fix it where it lives. A line in a shared skill helps every session after it.Check where the spend went, and how much of it was off target.Usage isn’t the goal. Shipped work and fewer redos are.
Before you go
What we’re learning, summarized.Big model for planning and hard debugging, mid-size for building, small for lookups.Live in the terminal? Try the desktop apps for a week.The jumps keep coming, and the gaps between them keep shrinking.The next twelve months will move faster. This is the slowest it will ever be.Take that task from the start and hand over the whole job around it. Then read the transcript back.Thanks for reading.
