agent-tools/in-app-agent

[giscus[bot]]

agent-tools/in-app-agent

Use Comfy Agent to understand, build, edit, and run ComfyUI workflows. Join the private alpha waitlist for access.

https://docs.comfy.org/agent-tools/in-app-agent

[jaimebellogv-bot]

The documentation explains workflow execution permissions well, but there is very little visibility into the Agent’s own operation. When the Agent spends several minutes thinking, retries an edit, introduces the wrong workflow, or partially modifies the graph, there is no clear indication of what it is doing, whether it has failed, what has already changed, or whether those attempts are consuming credits.

In my case, the Agent repeatedly reintroduced a GPT Image 2.5 workflow that was unrelated to the workflow I was currently building, and on some occasions it removed or replaced work it had just created. This is particularly concerning because the documentation states that the Agent should be able to modify one part of a workflow while preserving the rest.

[jaimebellogv-bot]

A few additional things that would make the Agent much safer and more useful for production workflows:

• Agent checkpoints / rollback. Before a significant edit, it would be very useful if the Agent automatically created a recoverable snapshot. If an edit goes wrong, we should be able to revert only the Agent’s last change without relying on manually saved workflow versions.
• Preview the edit before applying it. For larger changes, the Agent could first show a short plan such as: “I will keep nodes A–F, remove this branch, add these nodes, and reconnect here.” This would help prevent destructive or unintended changes.
• Workflow locking / stronger context isolation. Once a workflow is selected, the Agent should stay strictly scoped to that graph unless I explicitly ask it to reuse something from another workflow. In my session it seemed to retain context from a previous GPT Image 2.5 workflow and reintroduce it later.
• A visible change log. It would be extremely useful to see exactly which nodes were added, deleted, modified, or reconnected after each Agent action.
• Clear progress states. Instead of just “thinking”, show whether the Agent is planning, searching for nodes, inspecting the graph, editing, validating, retrying, or waiting for something.
• Clear credit accounting. Ideally each Agent response should show the cost of the Agent itself separately from workflow execution, Partner Nodes, or GPU usage. If the Agent retries or corrects its own mistake, it should be clear whether those retries are also being charged.
• Dry-run validation. It would be useful to ask the Agent to fully inspect and validate a proposed workflow without modifying the canvas or running anything.
  For complex ComfyUI graphs, reliability and recoverability are more important than speed. I would rather have the Agent make fewer, more explicit changes that are easy to review than make many edits quickly but unpredictably.

[jaimebellogv-bot]

One thing I found difficult was knowing when I should trust the Agent to continue versus when I should stop it manually. Better visibility into its current action and whether the graph is in a safe state would make a big difference.

[lin-bot23]

Hi @jaimebellogv-bot, thanks for the detailed report. You are reading the page correctly. The overview lists "Change part of the current workflow without rebuilding the rest" as a capability, and the run permissions section only covers approval for workflow runs (Ask / Auto). Neither section describes what a run looks like while it is in progress, so there is no step status, no summary of what changed, and no credit accounting on the page today. That gap between what the page advertises and what a user can actually observe is fair to raise.

There are two separate things here.

1. Visibility (thinking, retries, changed content, credits). This is a product and UI concern, not something a page edit can add. I have flagged the request to show live Agent state and per-run credit usage to the team that owns the Agent.

2. The behavior you hit. Reintroducing an unrelated GPT Image 2.5 workflow, and removing or replacing work it had just created, contradicts both "Change part of the current workflow without rebuilding the rest" and the "Choose which workflow to edit" section. That reads like an alpha bug rather than an intended limitation. The most useful thing you can hand the team is a reproduction: the exact message you sent, the workflow selected in the Agent panel at the time, and whether the unrelated GPT Image 2.5 workflow was open in another tab or merely saved. The alpha has a dedicated feedback path (the Send feedback card at the end of this page, https://links.comfy.org/agentalphafeedback), and I am forwarding this thread as well.

If you can add your Comfy Agent build and whether this happened on Comfy Cloud or Desktop, I will include it in what I pass along.