glossary · Running it

Attended execution

definition

Attended execution is the mode where a human drives the AI session while a workflow engine supplies the process: the agent asks for the current step, does the work with the human present, and advances.

in one line: You drive the AI. The engine supplies the steps.

in context

Where does attended execution fit?

You drive. The process rides along. The AI asks for the current step, does it with you, and moves on.

you askin your AI tool
get the stepfrom ConvOps
do the workwith you there
your OKapproval
next stepadvance
two ways to say it

What is attended execution, in plain words?

Same idea at two depths: the plain version, then what the engine actually does.

This is how most AI work happens today: a person in a chat. Attended execution keeps that, and adds the same steps and stops every time.

There is nothing to install or run. Your AI tool is the runtime.

for engineers

The agent calls workflow_advance over MCP, receives the step, does the work and advances.

No executor, no dispatch, no backend. Those exist only for runs with nobody at the keyboard.

the longer answer

Why does attended execution matter?

This is the mode most real work happens in today: a person in Claude Code, Cursor, or ChatGPT, working with the AI. Attended execution keeps that experience and adds the process: the same task run by different people, or the same person on different days, follows the same steps and stops at the same gates.

Its most underrated property is cost of adoption: attended execution needs no execution infrastructure at all. If the engine is reachable (in ConvOps, over MCP), the client you already use is the runtime.

It is also the on-ramp to autonomy: a workflow you have run attended twenty times is a workflow you can schedule with confidence, because the process, not the person, was carrying the quality.

Attended execution is often dismissed as the manual mode on the way to automation. In practice it is where most of the value starts: the person brings judgment and context, the process brings consistency, and the AI does the heavy lifting in between. It is also the mode where gates cost least, because the person who would approve is usually already in the session. Many teams run most of their work this way and keep unattended runs for the few processes that have earned it.

in convops

How does attended execution work in ConvOps?

In ConvOps, attended execution needs only the MCP connection. Clients verified to work include Claude Code, Claude on desktop and web including Cowork, OpenAI Codex, ChatGPT, Cursor and OpenCode. The agent calls workflow_advance, receives the current step, does the work with you, and advances. Close the session and the run stays at its step; open another client tomorrow and it continues from there. Nothing is installed besides the server entry, and no executor is involved.

questions

What do people ask about attended execution?

What is the difference between attended and unattended execution?

Attended execution runs in a client a person is using, with the person there. Unattended execution runs with nobody watching, started by a schedule or a dispatch to an executor. In ConvOps both follow the same workflow, the same gates and the same record.

Do I need servers or runners for attended execution?

No. In ConvOps, attended execution needs only the MCP connection in the client you already use, such as Claude Code, Cursor or ChatGPT. Your client is the runtime and keeps its own model and subscription. No executor is involved.

Can I stop an attended run and continue later?

Yes. Every step is persisted, so closing the session leaves the run at its current step. Opening any connected client later, even a different one, continues the run from there.