Product updates — November 2–8, 2026

Control what Forward Deployed Agents can do. Organization owners and admins can decide what Claude, Codex and other agents connected through the FDA MCP may read, change, run or create, for everyone or for a group, and in each phase. A thunk's owner or an admin can limit agents on that one thunk, and you can let your own agent work on every thunk in a phase for a set time. People are not affected: you can still do everything your role allows in the app. See Control what Forward Deployed Agents can do.

Set how much the AI reasons. Reasoning effort sets how much the AI reasons before it answers, for the whole thunk in Settings → AI Settings or for a single step in the step's AI Settings. More reasoning can help on hard tasks, but each step costs more and takes longer. See Move a thunk to a new AI model.

Turn input channels on or off. Inbound Requests now opens on a Channels tab that lists each way work can reach the thunk (Email, each webhook, API, Chat UI and AI), each with an Enabled / Disabled switch. A disabled channel takes no new work. See Inbound requests.

Check a step's inputs before it starts. Two settings can have AI read a step's input values before the step starts: one looks for instructions aimed at the AI, such as text typed into a form field that tries to override your instructions, and the other checks each value against its property's description. A step whose inputs fail a check waits for a person instead of running. See AI governance.

A reply when someone writes to a closed conversation. When someone messages a hosted Chat UI or conversational API conversation after its work item is closed, the thunk can send a short reply instead of running the workflow again. It is on by default; you can turn it off or change the wording on the conversation property. See Conversation properties.

The Builder MCP is now the Forward Deployed Agent (FDA). The Account menu, the Allowed Thunks and grant dialogs and the help center use the new name. The page where you approve Claude or Codex connecting to Thunk.AI now shows the thunk's picture and name, which app is asking, and what it will be allowed to do. In the Thunk.AI plugins the server is now called thunk-fda; after your plugin updates, sign in to it once more. See Building thunks with the Builder MCP: a guide for AI agents.

Forward Deployed Agents can do more. An agent can now pick an AI model for a step or a custom AI tool, or clear it to use the thunk's default; set a tool's availability condition; move a Workflow Plan step to a new position; and change one part of a custom tool, such as its description or one input, without resending the whole tool. When it builds a thunk, it names the work items after your business object, for example "Expense Receipt". It can also check whether a thunk is ready to go into production. See Checking deployment readiness with the FDA MCP: a guide for AI agents.

See how work flows through a thunk. A thunk's How it works document now shows how work moves through the thunk, with a diagram of its steps, tools and connections, and the diagram is also in the .docx download. Folder tabs on the home list also offer a Diagram view of a folder's workflows, the MCP thunks they use and the systems they connect to.

Rate any assistant message. Owners and admins can thumbs-up or thumbs-down an assistant message while looking at a work item's conversation, even when they cannot send in that thread, and see all ratings under Monitor → Feedback.

Custom tool history and caching at a glance. You can search a tool's Call history for calls whose input or output contains some text, and new calls appear as they happen without a refresh. Cost and time now sit under a Metrics menu, and the Custom-built tools list shows a Caching chip on AI and code tools that reuse identical lookups. See Tool call history and performance.

Steps that use several tools at once. A step no longer stops with an AI model error when the agent makes several tool calls at once and one of them returns an image or file. A tool that finishes after its step was canceled no longer breaks the next step, and retrying a tool that already finished no longer hangs the step.

Clearer reasons when something is refused. When a change to a thunk's settings or members is refused, Thunk.AI shows the reason instead of a generic error. Deleting a column waits for the delete and says why it failed, and adding a column can no longer create a duplicate name.

Shared form links work without an account. Anyone can open a thunk's shared form link without signing in to Thunk.AI.

File previews and content folders. Spreadsheets preview inline, and markdown, text and JSON files display in place. When you copy a thunk with its content folders, the files are copied into the new thunk, and a file no longer gets stuck showing "Removing".

Safer links. A web-link value that points to another site without http or https, or that could run script, now shows as plain text instead of a clickable link.

Helicone is no longer a sub-processor. AI model requests now go through a proxy that Thunk.AI runs itself.