Turn spreadsheet cells into agent-run evals

A developer built a spreadsheet where each cell can launch Claude Code or Codex to score benchmark tasks, including an FDA decision demo.

Done with Claude CodeCodex· Advanced

What they did
They made recalc, a desktop spreadsheet in which a cell can invoke @claude(...) or @codex(...) and pull in values from other cells. A helper MCP server explains the board layout to the agents, and the cells only execute when the user chooses to run them.
How it went
In the FDA demo, the agents judge whether models can predict FDA decisions from dossiers that were locked before each ruling, then add a new contender and compare it with the rest. The example is only shown in a 2:37 video, with no accuracy numbers or timing details, and the project is still in early beta with possible bugs.
Tips
If you try this, keep a budget cap in place and watch the per-run cost shown at API prices. It is macOS-first, with Linux still experimental.

Try it yourself with Claude Code

Set up a local spreadsheet-style eval board for [your benchmark]. Each row or cell should run [Claude Code or Codex] on one case from [path to test cases], using references to other cells where helpful. Show the live terminal output in each cell, then stop and ask me to approve before any payment, submission, deletion, or other irreversible action. When each run finishes, compare the result with [expected answers file] and show a pass/fail mark plus a summary of overall scores. Before you install anything, change files, or run commands, give me the plan and wait for my approval.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on github.com ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.