Windmill Adds Agent Settings and Saved Drafts

A Release Focused on Managing Agents

Windmill released version 1.804.0 on September 5, bringing two connected changes to its AI tools: a consolidated assistant settings window and a more durable way to edit saved agents. The update is aimed at the work around an agent, not at announcing a new language model. It makes configuration easier to inspect and moves saved-agent edits into the platform's existing draft-and-deploy workflow.

For teams maintaining reusable agents, the practical question is whether an edit changes a shared resource or merely leaves a separate configuration inside one workflow step. Windmill's new editor keeps the saved-agent link while storing changes as a resource draft. Those changes can survive a page reload and appear in Compare & Deploy.

There is also a migration detail worth checking before upgrading. Agents without an effective streaming setting now use streaming behavior by default. That affects recorded results and event storage, not just how text appears on screen. The release should therefore be treated as a workflow update with compatibility checks, rather than a cosmetic rearrangement.

One Place to Inspect Session Settings

The assistant settings window replaces the former read-only context panel. Windmill describes five areas: tools, skills, instructions, MCP connections, and linked files and folders. A control beside the model selector opens the window, bringing together settings that previously lived in different parts of the interface.

The tools area shows the definitions sent by the chat loop, including their descriptions and argument schemas. It is an inspection surface: the available tools still follow the session mode and connected services. Skills have individual switches and an editor, while workspace and user instructions occupy separate tabs.

MCP connections have both a reversible switch and a Delete action. Those should not be confused: the latter deletes the resource. Files and folders can show why content is not readable, such as pending indexing or missing access. These details help explain what a session can actually use, although the new window is not evidence that the underlying model has become more accurate.

Saved-Agent Editing Becomes a Draft Workflow

The agent form is now organized around Model, Messages, Tools, and Output. Core fields remain visible; optional fields can be added when needed. Windmill also keeps an emptied field visible while it is being edited, avoiding a disappearing row midway through a change.

The same grouped editor is used across the standalone resource view and relevant flow editing surfaces. Evaluations sit inside the editor, and a tool opens alongside the agent in a drawer. That places configuration, testing, and tool inspection closer together without requiring users to treat them as separate agents.

The larger change is where edits live. A saved agent stays linked to its workflow step, while edits are autosaved as a resource draft. Reloading the page does not discard that draft. Compare & Deploy can then show the change through the existing resource deployment mechanism.

This distinction matters when reviewing a shared agent. A draft is not the same as a deployed revision, and a successful save does not prove the agent performs correctly. Teams still need to examine the change and test representative inputs before deploying it. The release describes improved editing behavior, not an automatic approval process or a benchmarked improvement in task success.

What the New Controls Cover

The main changes address different parts of the agent lifecycle. This comparison separates inspection from editing and deployment so that a settings switch is not mistaken for a new model capability.

Saved-agent tool inputs also have a boundary: the reusable resource offers AI-filled or static inputs. Flow-specific values are bound on the workflow step instead. This keeps a reusable agent from depending directly on the context of one particular flow. It is useful to check that distinction when moving from an inline configuration to a saved agent.

Check Streaming Before Upgrading

The upgrade note says agents with no effective streaming key now stream. This includes more than newly created agents: configurations delivered through command-line tools, synchronization, APIs, imports, or saved resources may also leave the setting unset.

For text output, provider requests already used streaming. The change concerns what Windmill records and returns. Job results can gain a wm_stream field containing an event log, including full tool-result output, and runs write streaming-event rows. Downstream consumers that expect a particular result shape should be tested against this behavior.

Where the previous behavior is required, the implementation note specifies setting streaming explicitly to false on the step. That is a configuration choice to review, not a blanket recommendation for every workflow. Teams should consider which consumers read the result and whether event output belongs in their normal data-handling process.

An editing session left open during the upgrade has a separate caveat. The old in-memory saved-agent editing state becomes an unlinked step with its configuration inlined. Windmill says the configuration is retained, but the step needs to be linked again manually. Finishing or recording an outstanding edit before upgrading can make that transition easier to review.