AI tools briefing:
from prompts to workflows.

The useful question has changed

A text box can hide a great deal of machinery. Three recent announcements make that machinery part of the product: tools that assemble creative workflows, models that answer constrained decision questions, and agents that call existing automations.

The announcements below are facts attributed to their publishers. The practical implications are our editorial assessment. No performance, reliability or savings claim has been independently tested for this article.

1 October 2026 / Creative workflows

Comfy Agent launches in Comfy Cloud

Comfy announced an agent that can construct, run and revise workflows inside ComfyUI. The launch post says it is available in Comfy Cloud, describes it as beta and places Desktop availability a few weeks in the future. Desktop support is therefore an announced plan at the time of this briefing, not a released feature we can confirm. Read Comfy’s announcement.

What this could change: a creator may be able to begin with a description of the intended result and inspect a workflow assembled around it. Our interest is in whether that makes the process easier to understand and revise. A finished picture alone cannot answer that.

What to check: whether the workflow uses your intended models, whether intermediate results make sense and whether you can return to the saved work. Ask for an explanation of the steps before asking for twenty more variants. Quantity is a remarkably efficient way of postponing a decision.

29 September 2026 / Local model tooling

Ollama adds support for decision models

Ollama announced support for Jev-style decision models in version 0.35. Its release describes an API for asking named questions about supplied text and returning structured decisions, with ticket triage and routing among the examples. Three models are listed in the announcement; two are explicitly labelled experimental. Read Ollama’s release details.

Why the distinction matters: some jobs need a choice from a known set, rather than an eloquent explanation. For a proposed workflow, first write down the categories and what should happen when none fits. Otherwise the model will be beautifully organised around an ambiguous task.

A sensible first evaluation: use previously classified, approved examples and keep a separate group the system has not been tuned against. Review incorrect decisions by category. A fast answer that sends the difficult cases to the wrong place is still creating work.

This is an evaluation suggestion, not a benchmark result. The release’s performance figures come from specified tests and hardware; we are not treating them as a speed promise for your computer.

25 September 2026 / Business automation

n8n introduces a new Agents feature

n8n’s announcement describes agents configured with instructions, a model and permitted tools, including existing workflows. It distinguishes this feature from the existing AI Agent node, which the company says remains unchanged. The article also describes session visibility and approval controls. Read n8n’s announcement.

Our reading: the interesting design choice is the relationship between a flexible task and a predictable operation. A team may want an agent to work out what information is needed while keeping the actual update to a business system inside a tightly defined workflow.

What to test: choose a narrow task with an observable result. Examine which tool was called, what information it received and how a person can intervene. Include a case where the correct action is to ask for clarification or do nothing.

Check current availability and deployment requirements before planning around the feature. The general choice between managed cloud and self-hosting still has operational consequences; n8n’s deployment guide explains those options.

What connects these announcements?

Our assessment is that the workflow is becoming the useful unit of comparison. A model is only one part of it. Inputs must arrive in the right form, decisions need a destination, and somebody must be able to inspect what happened when the result is wrong.

That changes how I would evaluate a new tool. I would bring an ordinary example, an awkward example and an example that should not proceed. Then I would inspect the saved work, not merely applaud the first output.

There is room here for genuine convenience. There is also room for moving complexity behind another agreeable sentence. The difference becomes visible when you have to change one step, recover an earlier version or explain a result to a colleague.

Keep your next trial small enough to understand. A system that can do more should make its work easier to inspect, not make you feel impolite for asking.