This week you can organize your thunks into folders and create every kind of thunk from one New button, MCP interfaces are written in Thunk property types, and a new Behavior tab shows how each step of your workflow runs. AI coding agents that use the Builder MCP can do more without timing out, and custom code tools got clear limits and more reliable errors.
New
Organize your thunks into folders. The home page now has a Folders section: Owned for your thunks and Shared for thunks shared with you, with a breadcrumb and the folders inside the current one as large tiles above the list. To file a thunk, open its ⋮ menu and select Move (available to the thunk's owner and admins); a folder exists as soon as a thunk is in it. To move several thunks at once, or delete the ones you own, select their checkboxes and use the bar above the list. A thunk you create while a folder is open is filed in that folder, and a copy of your own thunk stays in the original's folder. A search in a folder also finds thunks in the folders below it. AI coding agents that build thunks through the Builder MCP can file, move and list thunks by folder, and keep each project's thunks together in one folder. See How to organize your thunks into folders.
One New button for every kind of thunk. The home page has a single New button at the top of its left menu. It asks what you want to create, a workflow, a chat app or an MCP server, and then walks you through the same steps for each: the environment (when you have more than one), the goal with templates for that kind, how the work goes, and the connections to use. Workflows and chat apps both ask for the process the work follows, and whether it runs on its own or needs a person's approval at some steps, and who approves. See Reuse thunk interfaces and plug-in connections for creating an MCP server from an interface file.
Write and read MCP interfaces in Thunk property types. An interface's tool inputs and outputs are now described in the same property types as a tool's inputs: Text, Number, Date, File, JSON Object, a Structure Type, or a list of any of them. View interface shows them that way, with Show JSON for the file. On a plug-in connection, Define interface… and Edit interface… open the same property editor a custom tool uses, and suggest the next version from what you changed. Matching compares types as property types, so a connection that describes the same types in its own way, including a third-party MCP server, now matches. A plug-in binding that satisfies its interface shows a simple check mark, and an Add case button adds a test case. See Reuse thunk interfaces and plug-in connections.
See how each step behaves, and which work items were slow for their size. Thunk admins have a new Behavior tab in Monitor. Click Run analysis to see, for each step, which tools its runs called and how often, how consistently runs followed the same sequence of tools, and which runs stood out because a tool returned an error, they used an unusual number of tools, or their sequence appeared in no other run. Your 10 most recent analyses are kept, so you can compare a step before and after you change it. On the Time tab, when Measure is Working time, a new Plot option, Per 100 output tokens, divides each work item's working time by how much the AI wrote for it, so items that were only long line up with the rest and the ones that were slow for their size stand out. See Behavior: reviewing how each step ran and Time Analysis: how long your thunk takes.
See uploaded spreadsheets without downloading them. An Excel (
.xlsxor.xls) or CSV file you uploaded to a thunk, such as a document in a content folder, now shows in place as a table, the way a Google Sheet already did. A workbook with more than one sheet has a picker to switch between them. Excel files in OneDrive and SharePoint show the same way. Each sheet shows up to 500 rows and 50 columns and says when there is more; download the file to see all of it. A file over 5 MB still shows as a link.
Improvements
A simpler home page list. One toolbar above the list holds the title (or folder breadcrumb), your search and every filter in force, with Search, Sort and Filter buttons and the card/table switch. Filters, search and sorting now cover all of your thunks, not just the ones loaded so far. Type and Phase are ordinary columns you filter on, in place of the separate By Phase and By Type menu entries; old links still open the matching filter. The table view has a new Created column you can sort and filter by, and Last modified starts hidden (turn it on from the column menu). An empty list says No thunks. The help, community and account settings icons have moved from the top bar into the account menu, and the left menu opens sooner when you hover over it. See How to organize your thunks into folders.
Custom tools: more room, clear limits and better inputs.
On a custom tool's Implementation tab, Inputs, Outputs, Details, Tools and Behavior are sections you can open and close; only Details starts open, and each closed section shows a short summary. A code tool's code now uses most of the pane. On the Definition tab, the Enabled switch has moved to the end.
One run of a code tool's script can make at most 50 tool calls and use about 4 minutes of CPU time, and Stop on a step now also stops a script that is still running. A script that goes over a limit ends with an error that says which one.
Ask the AI that builds a custom tool for an input or output that is a record with fields, such as an order line, and it can make it a JSON object from an example or use one of the thunk's Structure Types, as one value or a list.
Code tool scripts written by AI are formatted and commented. A saved script that is hard to read is still saved, and the AI is asked to send it again formatted. Scripts written before this change are not changed.
Try It on a thunk's tools is now shown only to the thunk's owners and admins, because a tool run uses the owner's connections.
See Custom tools and How custom code tools work.
AI coding agents can do more through the Builder MCP.
Set Required to start and Required to finish on a step's input and output properties. Reading a thunk's design now gives one list of each property's workflow input, workflow output and required settings, and what each step needs to start and must record to finish, so agents no longer confuse a workflow input with a required property.
Build complete custom tools: what each tool returns, its batch settings, and which tools an AI or code tool may call. An AI Classifier built by an agent can now be used. A tool built to match an interface can add output fields but cannot drop one the interface declares.
See each thunk's phase and owner when listing thunks.
When an admin who doesn't own a thunk runs one of its custom tools, the tool uses the thunk owner's connections, as Try It does, so tools that use the owner's connections no longer fail with "Tool … not found".
Choose a Compact Builder MCP on Account → Builder MCP → Access: fewer, smaller commands such as
steps.addin place of the largebatch_tools. After you switch, reconnect your AI app or refresh its tool list.When a Builder session starts, the agent says what it is doing, such as building a thunk, analyzing performance or debugging, and gets the matching guide. Your account's Builder MCP Sessions list shows the agent that ran each session; click a session to see its tool calls. API Keys, Allowed Thunks and your account's Connections are now tables with the Add button in the bar below them.
Trying a tool (
run_tool), running a Design Review (run_review), and creating or copying a thunk (create_thunk,copy_thunk) now reply within 20 seconds instead of holding the call open, which the network could cut off. A longer job returns a task the agent checks on withget_tool_run,get_revieworget_thunk. Agents built against the old replies need small changes: a running job now reportsworking,run_toolno longer takeswaitand its handle istaskId, and a review reportsworking,completedorfailedwith its findings inresult. If a new thunk is still being set up, the properties and steps asked for increate_thunkare not added; the reply says so.
See Building thunks with the Builder MCP: a guide for AI agents and "Choosing which outputs are required" in Schematized State.
The Design Review checks more of how instructions are written. It now reports, as advisory findings, step and custom AI tool instructions that don't open with what the work is for and follow it with a numbered process, instructions that recap a step's inputs or say "Do not change" a property, and instructions that define what a property's values mean. To keep a step from changing a property, bind it as an input and not as an output; what each value means belongs in the property's description. Rewrite applies the same guidelines. See How to Write Effective AI Instructions.
Agents can create a Google Doc or Sheet in a folder you name. When an agent creates a new Google Doc or Google Sheet, it can now put the file straight into a folder: give the folder's link in the step's directions. If you name no folder, new files go where they did before, to the top of My Drive.
Fixes
Custom and connection tools.
When a code tool's script calls another tool and that tool fails,
awaitnow always throws anErrorwhose message is the tool's error, such as Customer not found. Before, it returned the error as text starting with Code execution failed:. If a script checks a tool's result for error text, catch the error withtry/catchinstead.The Convert URLs to signed download URLs setting now works for tools from an MCP server or Generic API connection: the tool gets a download link that expires after 1 hour, instead of a link that needs a sign-in. A tool that uploads a file to another system could store an error page instead of the file, so if you turned the setting on for such a tool, check the files it uploaded before this fix. See Pass a file to a connected tool with a signed download URL.
A conversation link passed to a custom tool now uses the web address you sign in to, instead of always app.thunk.ai.
Date, Calendar Date, URL, Image, Location, Address and Country inputs and outputs added by the tool builder keep their type instead of becoming plain text, and a list it marks required stays required.
Workflow runs.
A copy of a Chat App keeps each step's Require Human to review and finish step setting as the original has it. Before, the copy's last step always required a person. If a copy you made earlier stops at its last step, open the step's Human-In-The-Loop Settings and turn the setting off. See Set human approval gates for your workflow.
A Yes/No output marked Required to finish must now be set before its step finishes. An output set to No counts as filled in.
Steps on muse-spark-1.3 stop less often while saving their results: Thunk.AI now repairs its malformed calls when what they save is clear.
Tests.
Suggest from this row suggests at most 20 assertions, aimed at the properties that best capture the result, and skips properties the step never filled in. An assertion that a property is empty now passes when it is.
A conversation test is graded after the agent answers the simulated user's last message, so assertions on what the agent recorded in that answer no longer show Failed by mistake.
A work item an AI coding agent turns into a test shows Treat this work item as a test turned on, and each of its conversations gets a simulated user that replays the recorded conversation.
JSON object fields stay in the order you defined. A property that holds JSON objects showed its fields sorted by name length, in the Table view of a list, in a list's Format views and filters, and in the schema editor. They now follow the order of the schema's
requiredlist, and the columns no longer move when you click a cell to edit it. See Format a list property.Smaller fixes.
A step shows its new directions when you return to the thunk's tab, or when the connection comes back, after an AI coding agent changed them.
Clicking inside a file's View content window no longer closes it, and clicking a file's preview no longer opens its editor; click the file's link below the preview to change the file.
The tools for a spreadsheet in a content folder are named from the sheet tab, without the file extension.
Through the Builder MCP, searching thunks ignores upper and lower case, and an agent can no longer list a built-in property such as Name or Status as a step's output.