> ## Documentation Index
> Fetch the complete documentation index at: https://docs.neoagent.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Iteration mode: step through an agent with Neo Support

> Run an agent in slow motion — it pauses after every tool output, Neo Support inspects each step with you, and you fix the instructions and rerun without leaving the chat

Tuning an agent used to be open-loop: edit the instructions, wait for the next run, read the
trace, guess, repeat. **Iteration mode closes the loop.** Tell Neo Support to run an agent in
iteration mode and the run pauses after **every tool output** — before the agent acts on it —
so you can watch the agent think in slow motion, catch the exact step where it goes wrong, fix
the instructions on the spot, and rerun from the beginning. All inside one chat conversation.
(Routine plumbing — like the agent loading one of its skills — doesn't checkpoint; only steps
worth reviewing do.)

## Starting a session

In any Neo Support chat:

> "Let's run my Daily Ticket Report agent in iteration mode."

Neo Support confirms the target (for a triggered agent it asks which ticket to run against),
freezes the agent's real schedule so a genuine run can't collide, and starts the run. There's
nothing to configure — iteration mode *is* the configuration: pause after every tool output.

Iteration mode works for **triggered and scheduled agents** (the ones that run on tickets or a
schedule). Chat agents and Neo Support itself can't be iterated.

## The rhythm

1. **The agent takes one step and stops.** The tool runs to completion; its output is captured
   but not yet shown to the agent.
2. **Neo Support inspects the output and narrates it** in a line or two: *"Checkpoint 1 —
   searched Autotask, 47 tickets back. Looks right for yesterday's window. Continuing."*
3. **On track → continue.** By default Neo Support continues on its own when a step is clearly
   right, and only stops to ask when something looks off, when the next step would be visible
   to the outside world (an email, a ticket note), or when you asked it to check something. Say
   *"check with me before every step"* to approve each one yourself.
4. **A step comes back wrong → fix and rerun.** Neo Support diagnoses the instruction that
   caused it, proposes the exact edit, applies it once you agree, and reruns the agent **from
   the beginning** against the fixed instructions — the fix has to prove itself from a cold
   start, the way the real schedule will run it.
5. **Wrap-up.** When a run comes through clean, Neo Support recaps every instruction edit made
   during the session (each one revertable), finishes the session, and unfreezes the schedule.

## What you see in the chat

While a session is running, a **banner sits above the message box** — *"Iterating «your
agent» — attempt 2 · auto-continuing verified steps"* — so it's always obvious the agent is
frozen mid-run and being supervised from this conversation. Each pause also drops a small
**"Iteration checkpoint" marker** into the thread between Neo Support's narration; click one to
see the raw checkpoint details. The banner disappears on its own when the session finishes.

## Billing and limits

* A session costs **one normal run per attempt** (the initial run and each rerun). The
  per-step pauses and continues are free.
* One iteration session per agent at a time.
* A session you walk away from stays parked — Neo Support nudges you once after \~30 minutes,
  and the session expires after 72 hours (the agent's schedule unfreezes automatically). A held
  step is never auto-continued.

## Good to know

* Instruction edits made during a session are real: they apply to the agent immediately,
  including its next scheduled run. That's the point — but it's also why the wrap-up lists
  every edit with the option to revert.
* Tool calls that error don't pause; the agent sees the error immediately and self-corrects,
  and Neo Support mentions repeated failures in its narration.
* Writes that normally require [technician approval](/core/technician-in-the-loop) still do —
  approval happens in the same chat.