Deploy an agent and walk away
Autonomous agents, built from templates, take a task and work it end to end — using your Brain for context — then report back. Pick one up whenever you’re ready.
“Deploy an agent to review my pull request and summarize the findings.”
- Start from a template, point it at the work
- It runs on its own and reports back
- Uses your Brain so it already knows your context
- Run more in parallel as you move up tiers
Chat is a conversation: you ask, it answers, you stay in the loop for every turn. That is the right shape for thinking and the wrong shape for work that takes twenty minutes of reading and produces one paragraph you actually need.
An agent is the other shape. You describe the job once, deploy it, and leave. It works the task end to end and leaves a report. The interesting part is not that it is autonomous — it is that you get to stop supervising, which is the only way handing off work is actually a saving.
Start from a template
You do not write an agent from scratch. You pick a template that already knows the shape of the job — review this, research that, triage this pile — and point it at the material. The description you give it is a sentence, not a specification.
Bounded work with a clear finish line is what this is good at. "Review my pull request and summarise the findings" has an obvious end state and an obvious deliverable, which is why it works; "manage my inbox forever" does not, and no amount of prompt writing changes that.
- Code review — read a diff, flag what matters, write it up
- Research — go through a topic and come back with something you can use
- Triage — sort a backlog or a pile of material into an order worth acting on
- Summarising — long material in, the paragraph you needed out
It already knows your context
An agent that starts from zero spends its first half re-learning things you already told the assistant. These read from your Brain, so the project, the stack, the constraints and the decisions you made last month are in scope from the first step rather than pasted in as a preamble.
The result comes back as a written report you open when you surface. Nothing is applied on your behalf: an agent reads, reasons and writes, and every decision about what to do with its findings stays with you.
How many at once
- Free — your baseline agent limit, enough to see whether the shape suits you
- Pro — five times the agent limit
- Max — twenty-five times the limit, running in parallel
Parallelism is the thing that changes how you work. One agent is a helpful errand. Four running at once while you are doing something else is a different working day — you queue the reading, the research and the triage before lunch and read three reports after it.
Where agents run today
What can an agent actually touch?+
Material you point it at and the context in your Brain. It produces a written report. It does not have your bank, it does not send messages as you, and it does not ship anything.
How long does one take?+
Minutes for most jobs, depending on how much there is to read. You do not have to stay in the app — it keeps going and the report is waiting.
Can I run several at once?+
Yes, up to the limit on your plan. Free has a baseline, Pro five times that, Max twenty-five times and in parallel.
What happens if it gets something wrong?+
You get a report, not an action, so a wrong conclusion costs you a read rather than a mess to undo. That is the deliberate tradeoff at this stage.
Try it
Free to start, no card. iPhone is the most complete app today and the web app has most of it; Android and macOS are in development.
Start free