Your agent can't do math
Language models are famously bad at arithmetic. They miscount the r's in strawberry. They think 9.11 is bigger than 9.9. Everyone laughs, posts the screenshot, and then hands the same model something that actually matters: a forecast, a hiring plan, a runway number.
The embarrassment is treated as a bug that scale will fix. I think that's the wrong lens entirely. Humans are bad at arithmetic too. We transpose digits, we drop minus signs, we cannot multiply four-digit numbers in our heads. Nobody considers this a crisis, because about five hundred years ago we stopped doing arithmetic by hand and started building tools that do it for us. The most recent of those tools became the most successful programming language in history, and almost no one calls it one. A billion people who would never call themselves programmers write formulas every day. The spreadsheet didn't win in 1979 because it could compute. Computers already computed. It won because it let you see the model.
So the obvious move, when your agent needs to work with numbers, is to hand it a spreadsheet. Everyone is doing this right now. Excel plugins, Sheets integrations, agents that read cells and write cells. And it works badly, for a reason nobody stops to examine: every single convention in a spreadsheet is an accommodation to human eyes. Cells sit at coordinates because eyes navigate by position. A formula gets copied into four hundred cells because you can only read one cell at a time anyway. B14 means March payroll only because of what's printed above it and beside it. None of this is what calculation wants. All of it is what eyes need. We made the grid flat because our eyes are flat, and then spent forty-seven years forgetting the flatness was a compromise, not a fact.
An agent doesn't look at a spreadsheet. It reads one. And read as text, a spreadsheet is a catastrophe. The file floods the context window one address at a time, and every accommodation that helped eyes now actively hurts. We keep handing agents our prosthetics and wondering why they limp.
See the flood live on the front page
This has been tried once before, in the wrong century. In 1991, Lotus shipped Improv, which removed the grid: named items, categories, formulas written like sentences. It was right about everything except its reader. Humans, offered structure, chose the scratchpad, because seeing is how we think. Improv died, the grid won, and the question closed for thirty-five years. Then the reader changed.
So we built the spreadsheet that would have existed if no one had ever needed to see it. Meaning lives in names: payroll = headcount * salary, written once, not four hundred times. Dimensions are real axes, as many as the model needs. Month, team, scenario, region are places you address, not rows you defend. Your agent writes revenue[month=Mar, region=EMEA] and never learns which row March is. The interface is tools, because tools are how agents touch the world. And the arithmetic itself never happens inside the model. It happens in a calculation engine that is deterministic all the way down, where every number carries its lineage, every change can be reverted, and a wrong answer can be traced instead of debated. Your agent doesn't get better at math. It gets what we got: the ability to stop doing math at all, without ever being wrong.
What that unlocks is simple to say. Your agent can build you models now. A hiring plan with scenarios. A forecast that updates when the assumptions do. A scratch table with real math behind it, stood up in one tool call instead of a provisioned database. Ask how it got a number and it shows you the chain, because the chain is real, not reconstructed. The format we use to read existing spreadsheets is open and MIT licensed, with the benchmark public: the highest accuracy of eleven formats we tested, at a fraction of the tokens, the only top-tier format that never overflowed a context window in nineteen thousand model calls. The engine is not a demo. It runs real company finances in production today.
Spreadsheets became the floor of knowledge work because somebody built the right form factor at the right moment and everything got built on top of it. Agents don't have a floor yet. Somebody's engine is going to be what every agent reaches for the moment it has to calculate something, and whoever builds it will have built the VisiCalc of a species that doesn't have eyes.
You can just run your own numbers now. So can your agent.