Windmill’s October 28, 2026 release adds a way for teams using its automation platform to build, test and run reusable AI agents through chat. Internal-tool developers and operations teams can now develop an assistant and try it out within the same conversation.
An agent is an AI assistant that can use software tools as well as answer questions. Windmill already had reusable agents; this update brings authoring and execution into AI Sessions, its conversational building interface. The implementation notes say sessions previously could not run standalone agents.
A practical example in Windmill’s released agent guide is an assistant that searches product documentation. Teams can save its instructions, model settings and tools, then reuse that configuration in other workflows.
Access is not open to every employee. AI Sessions documentation says users whose roles cannot write code cannot use sessions. Its default Ask permission mode waits for confirmation before writing, running or deploying; Yolo mode accepts those confirmations automatically.
A draft test can reach real tools
Testing an agent draft does not automatically isolate the systems it can reach. Windmill’s versioned agent guidance says a tool pointing to a saved script runs that script’s deployed version; a script that exists only as a draft will not work as that tool.
The practical implication is to check the connections behind those tools before a trial. Windmill’s dev workspaces can hold separate resource and variable values, allowing teams to configure different connections for development and production. Enforcing a lock against direct production changes, however, requires Enterprise protection rulesets.
Shared agents also make changes travel further: linked workflows load an agent when each step runs, so an edit can affect steps that have not yet started. Before deploying a revised assistant, identify which workflows depend on it.
Free software is not free AI usage
Windmill currently lists its free self-hosted edition at US$0 a month, with unlimited executions, three workspaces and up to 50 users. That is a platform offer, not an allowance for hosting or AI model usage. Sessions use configured model-provider access, including supported local models.
There is also a distinction between the cost of building an assistant and running it. The AI usage dashboard covers sessions and side-panel chat, but excludes agent steps in workflows; some costs are estimates or remain unpriced. Its total should not be treated as an all-in cost for the assistant.
The change is included in version 1.851.0. AI Sessions exist on Windmill Cloud and self-hosted instances, but the release does not establish that every installation has received this update. Existing teams should confirm their version and access before planning a trial.
For the related question of using chat to propose changes to company software, see Replit’s workplace code-proposal workflow.