Agents

What runs by itself vs what needs a human

Agents are AI set up to carry out multi-step work on their own, and the key question is which work you can hand over and which still needs a human watching each step.

An agent is not just a chatbot you ask questions. It is an AI that can take a goal, break it into steps, and carry them out across tools and systems without someone managing each move. An agent might read new emails, classify them, draft replies, update a database, and send a summary, all without a person clicking anything in between.

Agents are genuinely powerful for work that is well-defined, repetitive, and low-risk to get slightly wrong. They are a bad fit for work that requires nuanced judgment, consequences that matter a lot, or situations where getting it wrong would be hard to detect. The honest test: if the agent makes a mistake on this task, how quickly would you notice, and how easy would it be to fix?

Most teams are not ready for full agent autonomy, and that is fine. The useful starting point is identifying a few tasks where an agent could handle the legwork and surface a result for a human to review, rather than acting on its own.

Make it visiblePick one recurring multi-step task and write out every step a human takes to complete it, including the judgment calls and edge cases. That description is the starting brief for any future agent, and the exercise often reveals why it is not ready to automate yet.

Why AI needs this

Each part of your work matters to AI in a specific way. Some of it is context a tool needs before it can help, some of it is work a tool can take on, and some of it is judgment that should stay with you.

Work without constant supervision

An agent can run a task end-to-end while you work on something else. It is not a replacement for human judgment but a way to remove the manual steps around it.

Clear handoff points

Good agent design includes explicit moments where a human reviews before the agent proceeds. AI needs those checkpoints built in, not added as an afterthought.

Well-defined inputs and outputs

Agents work best when the task has clear edges. AI cannot operate reliably on fuzzy briefs; the more precisely you can describe what done looks like, the better the agent performs.

Logging what ran

An agent that runs invisibly is an agent you cannot audit. AI processes need logs, so you can see what happened, catch errors, and improve over time.

Questions to explore

Use these on your own or in a group. There are no right answers, only better conversations.

Which tasks in your work follow a predictable enough pattern that an agent could handle the steps?
How would you know if an agent made an error? Is there a natural review point in the process?
Which parts of your work require context, relationships, or judgment that an agent could not have?
What is the worst realistic outcome if an agent gets a task partially wrong?
Who on the team would need to understand how an agent works in order to supervise it?

Readiness traps

Giving an agent access it does not need. An agent that can send emails, modify databases, and delete files is an agent where a small error becomes a big problem. Start with read-only or draft-only access.
Treating an agent as set-and-forget. Agents drift when the world around them changes. Regular review of what they are doing is not optional, it is part of how they stay useful.
Automating a task before you fully understand it. If you cannot describe every step a human takes to do this task, you cannot build a reliable agent for it.