This week you can set up Claude or Codex as your Forward Deployed Agent from the home page, and an AI agent connecting to Thunk.AI for the first time now gets you to a running thunk. A chat tool call opens straight into its Call history, and custom code tools show what their calls cost and how long they take. My Thunks updates live, copies of a thunk keep its environment, and pausing a thunk no longer reports errors.
New
Set up Claude or Codex as your Forward Deployed Agent. Under New on the home page, the Claude and Codex buttons show how to set up either one as your Forward Deployed Agent: an AI agent that builds, tests and improves your thunks. See Build thunks from Claude Code or Codex.
A guided first run for AI agents. When you connect Claude or Codex to Thunk.AI for the first time, the Builder session asks your agent to explain what a thunk is and to offer to copy the expense receipts sample, so you can see a running thunk. See First run with the Builder MCP.
Open a chat tool call in its Call history. From a tool call in a step's chat, you can now open that call in the tool's Call history. See Tool call history and performance.
See what a custom code tool's calls cost and how long they take. A custom code tool's Call history now shows the cost and time of its calls, including how much of each call was spent waiting on the tools it called. See How custom code tools work.
Filter work items by when they were created in the Builder API. A work item query through the Builder API can now filter on when each item was created. See Building thunks with the Builder MCP: a guide for AI agents.
Improvements
My Thunks updates live. My Thunks now opens on Owned, Shared or Samples, and thunks created or shared with you appear without a refresh. New conversations, connections and environments created elsewhere also appear without reopening the page. See How to organize your thunks into folders.
Building thunks from Claude and Codex.
Codex makes far fewer failed tool calls when it builds a thunk, including when it adds money fields.
When an agent sends a field the Builder MCP doesn't recognize, the error suggests the field it likely meant and lists the valid ones.
The Builder MCP now shows the output of document template and spreadsheet lookup tools.
Adding thunk members through the Builder MCP adds teammates from your organization directly, instead of asking someone to confirm each one on the Team page.
The Thunk.AI guidance plugin for Claude and Codex is now called thunk-skills, and its skill is thunk-guide. Claude installs move to the new name on their own; in Codex, remove the old plugin and add the new one.
Step timeouts say where the time went. When a step stops because the AI model did not answer in time, the error now says where the step's time went: earlier AI calls, tools, or the call itself. It no longer says the AI provider was slow or not responding when the step simply ran out of time, so you can tell a slow provider from a step doing too much work. See Troubleshooting errors.
The Monitor cost chart covers your whole window. The cost chart in Monitor spans the whole time window you picked and shows each rerun of a work item as its own point. See Token costs: what your thunk spends on AI.
Kimi K2 Thinking is retired. Google is retiring the model on Vertex. A thunk still set to it shows the model as unavailable until you choose another model, such as Kimi K3 or GPT-5.1. See Supported AI models and what happens when they change.
Fixes
Copying a thunk.
A copy now keeps the original's resource environment when you can access it, so it uses the same AI models. When the copy must use your personal environment instead, you are warned.
A copy or import now saves your default AI models when the original had none set.
A copy of a sample thunk now brings over the files in its content folders.
MCP servers.
A failed call to an MCP server tool is now treated as an error in your thunk's steps and code tools, instead of passing the error text back as a result.
Thunks connected to MCP servers hosted on Cloud Run no longer run out of server capacity from idle connections.
MCP connections to private or internal network addresses are now refused.
Workflow runs.
Pausing a thunk no longer reports its in-flight runs as internal errors or sends error-alert emails; they show "Stopped because the thunk was paused."
A step whose run condition, finish condition or required input uses the work item's Name now sees the name, so it runs, skips and starts as configured.
A tool input that refers to a Structure Type column no longer makes every run fail.
A chat app whose conversation column has a name other than the default now gets the full chat app behavior: its opening message, error messages and test replays. Setting up a chat app no longer turns on human review for its last step.
With Detect and block dangerous content in tool outputs on, a thunk's own policy search and work item updates are no longer blocked as malicious.
Dialogs and buttons. Delete and confirm dialogs stay open until the action finishes, and Save and Create buttons ignore extra clicks while they are working. Copying an API key, email address or webhook URL tells you when the browser blocked the copy, instead of saying "Copied". Clicking Recent changes on a thunk no longer reloads the whole page.
Keyboard and screen readers. Work item cards, the automated-check status icon on a work item, the deployment environment list, the page header's Help button, several step-settings switches, the assignee picker and step links can now be used with the keyboard and screen readers.