Every agent so far has waited for you. You write, it works, it stops, and then it waits for the next order. That is the right shape for a task and the wrong shape for a job. Nobody hires a marketer and then messages them every morning with what to do next.
Autonomous agents are proactive. Instead of you handing out task after task, you give an agent a goal and it works toward it on its own: it decides what matters most, does it, and comes back to you only for the calls you asked to make yourself.
How it works
Autonomy settings for a topic: the job, goals in priority order, what it must ask you first, and how often it checks in.
- Turn autonomy on for any topic from the Autonomy chip next to the model picker, or choose "On its own" when you create an agent
- Write the job, the goals, and how it should work with you, in plain English. Or let it draft them from the conversation you already had
- It drafts a plan first: recurring tasks and a prioritized backlog. Nothing happens until you sign it off
- On its schedule it checks in and works task after task, keeping the plan current as an artifact you can open any time
- One agent can be autonomous in several topics, each with its own job, goals, and schedule
Its first check-in: the agent drafts a plan from the job and goals, then waits for your sign-off before it touches anything.
Proactive, but it still asks
Working on its own does not mean acting without you. Your preferences decide what it does alone and what needs your approval. "Ask me before emailing anyone outside the company" means it drafts the email and shows you. When it needs you, it asks with a question card, the topic turns amber, and check-ins pause until you answer, so a new check-in never buries your decision.
You stay in charge: edit its job and rules, make it check in now, or turn it off. The settings and plan are kept.
It gets better the longer it runs
- When you approve something and pick "don't ask again for this kind of thing", it records that as a rule and stops asking next time
- Every session it writes what it learned into its own memory: your preferences, what worked, what you corrected
- Its plan evolves as it goes: finished work moves to done, new ideas join the backlog, and dead ends drop off
So week three asks less and delivers more than day one, without you rewriting its instructions.
Originally published on the JackHamr blog. JackHamr is an AI agent platform where agents plan, build, test, and review software end to end. Free to start.
Top comments (0)