Skip to main content

Product updates — November 9–15, 2026

What changed in Thunk from November 9 through November 15, 2026.

This week you can schedule a thunk's tests to run on their own, test a custom tool that holds a conversation, and see Mermaid diagrams drawn in place. Org admins can store a provider API key once and share it across models and bring their own keys for more AI providers, and MCP and REST API connections can send extra headers. Forward Deployed Agents follow your organization's rules on its thunks and can do more, and the compact work item view shows what each step did again.

New

  • Scheduled test runs. You can schedule a thunk's workflow and tool tests to run automatically and read each run's report in Test Results. In Testing → Scheduled runs you can filter which test work items a run includes. A test work item that is waiting on a person is graded right away instead of at the run's deadline, and the report shows the workflow status each item reached. See Automated tests for workflow steps.

  • Test a custom tool that takes a conversation. A custom tool's test case can now include a conversation and a simulated user who answers the tool's questions, and the test can check what was said. See Automated tests for custom tools.

  • Mermaid diagrams. Mermaid diagrams written in chat replies, documents and text values now display as diagrams instead of code.

  • Share one provider API key across models. Org admins can store a provider API key once as a credential, attach it to every model definition that uses it, and rotate it in one place. See Orgs, roles and environments: enterprise administration guide.

  • Extra headers on connections. MCP and REST API connections can send Additional Headers, such as a warehouse or tenant id, on every request. They are stored as a secret. Keep API keys, tokens and Authorization on the connection's authentication method instead. See Connecting to external systems and tools.

Improvements

  • Your own keys for more AI providers. Org admins can add their own API keys for Fireworks AI, Cohere, Meta and OpenRouter, and runs on those providers' models use the organization's key. When several of your keys can serve a model, the model's own provider key is used, and an OpenRouter key serves only models none of your other keys can. OpenAI models reached through OpenRouter can use the Responses API once the model has input and output prices set. See Understanding LLM costs and latency.

  • Forward Deployed Agent rules follow the organization's thunks. An organization's Forward Deployed Agent rules now apply to the organization's thunks, whoever's agent is acting. Owners and admins can see each person's effective agent access on a thunk in the Users and FDA tabs, and the approval page shows which Thunk.AI account you are signing the agent in as. See Control what Forward Deployed Agents can do.

  • Forward Deployed Agents can do more. An agent can wait for a work item run to finish instead of checking on it again and again, allow only a chosen list of tools on a workflow step, rename a work item, and download a generated PDF or Word file. When it looks up a custom tool's call history, each call comes with a link that opens it in Thunk.AI. See Building thunks with the Builder MCP: a guide for AI agents.

  • Go further back in history. A tool's Call history loads older calls with Load more, and every call has its own link that opens it however old it is. The Recent changes dialog can also load changes older than the newest 100. See Tool call history and performance.

  • Catch problems before they ship. The design review flags a code tool whose code does not parse, or that calls another tool by a name it cannot reach or with inputs that tool no longer has. The API tool editor warns about header lines that won't be sent. Before you move a thunk to another environment, the compatibility review also flags models on custom AI tools, the tool builder and test runs that the new environment doesn't offer.

Fixes

  • The compact work item view shows what each step did. Expanded steps show their property changes and messages again instead of "No changes or messages from this step". Each step also lists the tool calls that changed something, folded steps sum up on one line, and the Workflow header shows the work item's Status.

  • More dependable custom tools. Code tools can read files up to the documented 10 MiB with getBytes(), and reading a large file more than once no longer fails now and then. A code tool that waits for a reply continues after the person answers, and code tools can call MCP connection tools without passing every optional argument. Saving a custom tool no longer duplicates its unit tests, keeps removed fields, turns off a new tool's Enabled switch, or drops other tools when the page was out of date.

  • Long and paused steps. Long step conversations no longer drop messages after the first 256, and a step that paused to wait for a reply no longer fails on its next AI call. Steps on a Gemini model with your own Google AI Studio key on the Flex processing tier fall back to the Standard tier when Google reports high demand, instead of failing.

  • Clearer messages. When an AI provider rejects a key, the error says whose key it was and who can fix it. When a field or an admin setting can't be saved, Thunk.AI says why. Adding someone who already has a role in your organization or thunk offers to move them to the new role instead of showing an error.

  • Security fixes. Fixes from our annual penetration test, including stricter checks on where tools can send requests and on who is recorded as having last updated a value.

Compliance

  • Langfuse is now a sub-processor. Thunk.AI sends AI model requests and responses to Langfuse Cloud (United States) to trace and diagnose model calls; these can include workflow content. Langfuse, part of ClickHouse, processes this data under a data processing agreement and holds SOC 2 Type 2 and ISO 27001 certifications. Helicone was removed as a sub-processor last week.

Did this answer your question?