Week of September 13: Open the Spreadsheet

Open the Spreadsheet

Last week the workforce started before you arrived. This week it stopped stopping at the edge of the platform.

Almost everything below points at one claim: the work leaves Supanova and lands in the tool you already keep your work in.

What's New

The spreadsheet comes back with something in it

A task that produces an artifact now has to read its own work back before it can report success.

Creating the file is not enough. A header row is not enough. A write acknowledgement from the provider is not enough. The cells have to be there when the atom looks again.

If population cannot be verified, you are told the existing sheet needs filling, rather than being offered a brand new empty one.

It reuses the file it already made

Artifacts survive a retry now. A sheet created on the first attempt is found and reused on the second instead of being duplicated beside it, and the same goes for the folder it lives in.

Research already gathered is checkpointed against the task brief and carried forward, so a second attempt continues rather than paying for the same sources twice. Receipts survive the handoff into grounding review instead of being dropped on the way.

Your connected tools show up in the work

A task planned by the Project Manager was running with no connected tools at all, because the plan's author was an atom rather than a person, so it searched for that atom's connections and found none. It resolves the real requester now, and your workspace's connections reach the work that needs them.

Atoms can also chain actions across the tools they have. Make the folder, take the ID that came back, put the sheet inside it, fill it. Reversible operations like moving a file no longer demand full access, and a prerequisite that succeeded counts as progress instead of reading as a stall.

Gates got more honest in both directions. A brief that reads a repository and delivers a spreadsheet no longer asks permission to write to GitHub. Tools you explicitly excluded are not quietly turned into completion requirements. Provider failures wrapped in a success envelope are treated as failures.

When the same tool ran several times, the results collapse into one row with a call count, newest first, with the full history a tap away.

The model you picked writes the work

Assign a task to a model and that model writes the deliverable.

Research tasks used to be switched wholesale to a different, search-capable model. Now, when your model has no hosted search of its own, a helper gathers the source evidence and hands it back, and your model, with your reasoning settings and your connected tools, writes the result. The same holds for the corrective rewrite at the end.

Enabling a workspace tool no longer touches your models either. The model-migration dialog is gone, and the permissions save through the ordinary workspace flow regardless of which models you have assigned or locked.

GPT-6 Astra joins Contributors, Leaders, and Executives through the pickers you already use, with its costs and tool support registered like any other model. No new tier and no new entitlement.

A decision you can actually answer

A finished LinkedIn draft was being shown as a request for input, because its optional closing question contained words that read like a blocker. The preview underneath it said "I am Lisa Patel."

Optional follow-up questions no longer pause finished work. When a task genuinely needs a decision, it arrives as named choices with their consequences and a recommendation, and the option you pick keeps its identity across desktop, mobile, recurrence, and standalone views.

Persona boilerplate is stripped out of work previews. Older malformed pauses that cannot be repaired offer Continue with best judgment instead of pretending to be a question.

Talk over it and it stops talking

Chief of Staff voice and Voice Huddles moved to a native GPT-Live runtime.

You can interrupt it. While you talk, the thing being discussed is on screen: the briefing, the pending decisions, the plan, the alternatives. Edit the plan out loud and what you see stays in step with what you said. Open a deliverable, select a passage, and talk about that passage.

Each atom in a Huddle has a persisted speech profile, so the same atom sounds like itself every time. A Huddle with several atoms runs a separate session per active speaker and hands the floor between them.

Every change still goes through the exact proposal you were shown, at the revision you were shown. A general "yes" does not approve anything. A revision to a deliverable checks that the version under discussion is still current.

Voice usage is debited at what the provider charged, with nothing added on top, and recovery never invents elapsed time it cannot prove.

Confirm it and it starts

A project you and the Chief of Staff worked out by voice used to stop at a plan, even when the moment felt like starting.

Live voice now offers the two real outcomes plainly: Save as plan, or Confirm and start. Confirm and start creates the project through the Project Manager, waits until the charter, tasks, and dependencies are actually written, then puts it through the same approve-and-start path as anything else.

You can name the choice out loud, down to the alternative: "confirm and start the Friday option." If it cannot tell whether you described a project or a single task, both readings are shown and you pick one before anything is created.

Rate it, then clear it

Completed tasks carry thumbs up and down on the row itself, using the same review the task detail page uses.

Rate one and it leaves your queue, but only after both the rating and the removal are confirmed. A thumbs down keeps the revision form within reach. Removal is a personal preference and nothing else: the output, the costs, the activity record, and every deep link stay exactly as they were. Undo does not expire, and a Removed view holds everything you cleared.

One task, its own steps

A standalone task can carry subtasks now.

Ask for product context read, prospects researched, outreach drafted, and a spreadsheet delivered, and you get one parent task with the dependent steps it needs, rather than being told it has to become a project. Simple work stays a single task. The parent, the children, the dependency edges, and the context chain are written together or not at all.

Ask about a workflow and the workforce answers

Workflow setup questions used to go to Theo, which is customer support.

The button is Clarify setup now, and sending it creates a real task for your workforce with your answers attached, and gives you a link to it in Activities. The screen says plainly what that task will and will not do: it investigates and reports back, and it does not connect apps, change configuration, or unblock the workflow on its own.

A confirmed human step in a workflow no longer reports itself as a missing provider action, and Google Sheets destinations save properly now that their configuration lives in one record.

Under the Hood

Posting a comment could blank the discussion. The database was returning timestamps as date objects, one helper was turning them into a string that was not a valid time, and the browser threw while rendering the comment that had just saved correctly. The timestamp is normalized at the source now, the browser refuses to crash on a bad one, and a render failure inside a discussion stays inside that discussion instead of taking the page with it.

Deliverable revisions show their real state on the request that asked for them: queued, running, saved, or failed. The generation is durable, the version is written atomically, and a partial document is rejected before it is stored.

Atom records got their modification timestamp back, so capability learning and personality updates leave a trace of when they happened.

Sixty-seven weeks in. You built it, you talk to it, you trust it, and now the work lands where you keep it. See you next Sunday.