What caught my attention here isn’t the spreadsheet part. It’s the decision to make the spreadsheet the control surface for agent work instead of bolting agents onto an existing app and calling it a day. That feels more honest to how people actually use these tools: you want visible state, editable state, and a way to pin an agent down next to the evidence. Chat is fine until it isn’t. A grid is colder, but also more accountable.
The interesting bet is that a spreadsheet is the right primitive for human-agent collaboration. I think that’s plausible, but only for a pretty specific class of work. If you’re doing contract review, benchmark runs, tie-outs, or anything where “show me the source, show me the check, don’t let the agent declare victory on vibes” matters, the model makes sense. The source’s examples are strongest when there’s a clear audit trail: answers in one column, citations in another, a test or formula in the next cell over, and the whole thing turning red when reality disagrees. That’s a much better shape than a chat transcript with a few pasted snippets.
I’m less convinced by the “spreadsheet for everything” implication. There’s a whiff of universe-expanding here, and those projects usually hit a ceiling when the workflow stops looking like rows and columns. A spreadsheet is great when the task decomposes into repeated units and checkable outputs. It gets awkward when the work is exploratory, branching, or deeply stateful. The repo seems aware of that, which is good. It doesn’t pretend this is a replacement for a general IDE or a general agent runner; it’s a harness. That word matters.
The other part I like, and would actually want to try, is the refusal to hide the machinery. Cells can be values, formulas, terminal commands, actions, or agent calls. That sounds messy, but it’s the right kind of messy. If an agent is doing work that can be validated by tests, put the test in the same board. If it’s reading a document, put the quote next to the answer. If it needs to be steered, edit the cell instead of yanking on the chat thread. That’s not elegant in a product-marketing sense. It is elegant in a debugging sense.
The privacy angle also feels unusually grounded. No account, no server, no tracking, boards as folders on your machine. That’s not just a checkbox; it changes the whole trust model. Of course, the repo is also clear that this is not a sandbox, and that commands run with your permissions. Good. That’s the sentence I’d want prominent before anyone gets ideas about “safe agent automation.” Safety here is procedural, not magical.
My only real hesitation is whether people will tolerate the cognitive overhead. A board full of agent cells, check cells, review cells, and a “Needs You” list sounds powerful, but it also sounds like work. You’re trading one kind of hidden complexity for a much more visible kind. Sometimes that’s exactly what you want. Sometimes it’s just a fancier way to watch your own process become a spreadsheet. I’d be curious to see whether this becomes a niche power tool for people doing high-volume verification, or whether it turns into one of those beautifully clever systems that only makes sense when you’re already committed to the workflow.
