This week the work items list gets a Simple table that shows each item's state and progress, reports let you write your own instructions for the AI summary, and Thunk.AI suggests your next step in the Codex app. AI agents that build thunks through the Builder MCP can set up more of a thunk on their own. New thunks start with Draft mode off, steps save their results at the end by default, and editing the text on the Export page no longer publishes your tools.
New
A Simple table for work items and test rows. The Work Items and Test Rows lists now open in a Simple table that shows each item's state and how far it has got through the workflow. The Detailed table is one click away. Running work items show the same spinning square in both tables, and on a chat app the Simple table includes End User, so you can find a person's conversation. See How to work with your work items list.
Write the instructions for a report's AI summary. The report editor has a Summary instructions box. Say what the summary should cover, for example one line per region with its count. Leave it empty for a general summary. See How to save a report from your work items.
Next steps and thunk cards in the Codex app. After Codex works on a thunk, Thunk.AI suggests what to do next. When the agent starts or opens a thunk, a card for it appears on its own, with one-click Analyze, Explain, Debug errors and Suggest improvements. See Show a thunk in the Codex desktop app.
AI agents can set up more of a thunk through the Builder MCP. An agent can now set up a thunk's webhooks and email channel, and create an OAuth connection that you finish signing in to from a link to your Account page. It can also move a thunk to another environment, see each thunk's environment name, and set or backfill property values on existing work items. See Build thunks with the Builder MCP.
Participants can see the Tests pane. Participants, and anyone viewing a public thunk such as a sample, can now open the Tests pane read-only to see its test work items, assertions, reports and settings. On a test work item, the Out Result tab now shows the test assertions at the top, so you can see what passed or failed next to the values. See Evals: automated tests.
Improvements
New thunks start with Draft mode off. New requests to a new thunk start running right away. Thunks you already have keep their current setting.
Steps save their results at the end by default. A step now writes its results to the work item once, just before it finishes. A step set to stop on its first tool failure saves as it works, so its progress is kept, and you see a warning if you combine that with saving at the end. See Choose how a step handles errors and when it saves results.
Auto-finish inactive work items sooner. Auto-finish inactive work items after (minutes) now takes minutes, down to five. When a work item is auto-finished, its step chat says it ended because it was inactive, not that a user canceled it. See Data retention.
Builder MCP works better on large thunks. An AI agent can ask for only the steps it needs to read or change, and change part of a large code tool without sending all of its code again, so big workflows no longer overrun the agent's output limit. Builder MCP and Builder API arguments now use one name for each idea, for example
descriptionfor free text. A call that uses an old name is refused, and the error gives the new name. See Build thunks with the Builder MCP.Easier to read edits in version history. When a long field changes in only a few places, the version history hides the text that did not change, so you can see the edit.
Fixes
Editing the Export page no longer publishes your tools. Changing your thunk's name or description on the Export page now only saves the text. If the export is on, the new text goes to the live server. If it is off, it stays off until you turn it on. See Publish your thunk as an MCP server for AI agents.
Thunk cover images are made again. New thunks were getting a placeholder image instead of a generated cover.
Report summaries on large reports. A report's summary no longer fails with a bare "Error" on a large report. It summarizes the columns the report shows, and if something still goes wrong it says why.
All of a connection's tools can be used. When a connection has several tools with long names that begin the same way, the AI can use all of them again, not just the first.
Work item fixes. After a work item is canceled or auto-finished, its step chat stops spinning. A step that just started no longer appears twice in the step list, and filtering work items no longer shows an error when a saved filter names a column that is not there.