The AI page

What it is: an agent that runs along a path you’ve laid out, handling the judgement-heavy steps while the overall flow stays defined in code. The steps are known in advance.

Good for: processes where you already know the shape of the work and want it to run the same way every time. The most predictable kind, and usually the cheapest and easiest to make reliable.

What it is: you give the agent a defined job and it completes it from start to finish, such as gathering data, drafting a document, doing research or triaging incoming items. It works out the steps itself instead of following a fixed path.

Good for: multi-step work where the exact steps can’t be known up front. More flexible than a workflow agent but harder to make reliable, so it’s usually run with an approval step before anything critical happens.

What it is: an agent you interact with through conversation, in chat or increasingly by voice. It can answer questions, and if it’s connected to your systems it can carry out the request rather than just pointing at an article. The same idea works internally as agent-assist, sitting next to staff during a live conversation and pulling up information or drafting replies.

Good for: customer support and internal help, as long as it’s grounded in your own content so it doesn’t make things up. The agent-assist version is a lower-risk start, since a person still sends the final reply.

A coding agent writes and runs code to get things done. At one end that’s working in a codebase, writing, editing, fixing, and reviewing software with a developer approving what ships.

The more interesting end is in-product: instead of a fixed set of tools, the agent gets a sandbox and writes and runs code on the fly to handle problems you couldn’t anticipate, especially work over large or messy datasets. Code becomes the agent’s universal tool, run inside a constrained harness so it can do a lot without being able to do damage.

Good for: speeding up software delivery on well-defined work, and dynamic, data-heavy tasks where the request is different every time, like slicing a big dataset a new way for every question or making sense of whatever a user uploads.

What it is: several specialised agents working together on a larger job, where one breaks the work down and hands pieces to the others.

Good for: complex tasks that are too much for a single agent, or where different steps need different skills or different access. There are more moving parts, so this is used when the problem genuinely calls for it.