A personal agent is not a chatbot with your calendar

Connect a language model to your calendar, your mail and your files and you get a chatbot with better inputs. It is genuinely useful and it is not an agent, because the initiative still belongs to you: it waits, you ask, it answers. Agency starts where the waiting stops — and that turns out to require three things nobody ships in a chat window.

Standing goals rather than questions

You do not brief a capable assistant by asking a question. You give them something to hold: keep the pipeline warm, make sure nothing slips past its deadline, tell me if this supplier becomes a problem.

Standing goals are strange objects for a chat interface, because they have no reply. They persist, they get revised, they conflict with each other, and something has to decide which one wins on a Tuesday morning when both cannot be satisfied. That arbitration is most of what an agent actually does, and none of it looks like a conversation.

Its own schedule

The second requirement follows from the first. A goal that is only evaluated when you open an app is a to-do item, not a goal.

An agent has to run on its own clock: waking on a timer, on an event, on a threshold being crossed. Most of what it does in a week should be invisible — checking that things are still fine, and concluding that they are.

This is the part that makes people uneasy, and the discomfort is appropriate. Something acting while you are not looking needs to be bounded by rules that exist outside the model and logged in a way you can read afterwards. That is the price of the initiative, and a personal agent without it is either useless or reckless.

The judgement to stay quiet

Here is the failure mode nobody anticipates: an agent that works constantly and tells you about all of it.

Every notification spends something. Interrupt often enough with things that did not matter and the agent gets muted, which is the same outcome as it not working at all. So attention is a resource the agent has to manage, and the decision “is this worth interrupting a human for” is a genuine judgement — arguably harder than the work that produced the finding.

The rule we keep coming back to: an agent should be measured by how much it does without telling you, not by how much it reports.

What delegation actually needs

Put those together and the requirements are specific.

Boundaries that exist outside the conversation — what may be done silently, what needs confirmation, what is never allowed. Enforced by the system, not by an instruction the model might reinterpret.

Structured knowledge — an agent deciding on its own schedule cannot ask you what a commitment is every time. It needs a model of your world that persists.

Self-regulation — a way to notice its own accuracy dropping and narrow its own scope before you have to.

A record — everything it did while you were not watching, reconstructable afterwards.

That list is why a personal agent is an operating-system-shaped problem rather than an app feature. Three of those four are infrastructure, and none of them can be retrofitted onto a chat window.

The test

If the software only ever does something after you type, it is an assistant, and a good one is worth having.

If it can be given something to hold, act on it while you are asleep, decide that most of what it found was not worth waking you for, and show its work when you ask — that is an agent, and the design questions are completely different.