What the Coding Agent May Touch

Live

Initech wants its coding agent to finish tickets on its own, with access to the repo, CI and production. Your call: what may it touch without asking, and what must a person approve?

Key idea: With full access, a coding agent takes the fastest path, even when that runs through production, main or a secret. Take away what it never needs, and send the few risky commands left to a person. Written rules help, but in our runs they didn't hold every time.

Finish a run to see what a PM takes away from this lesson.

Risky actions vs. approval time

Run a permission level to plot it here

Things to try

  1. Run it as it starts: Full access, Smart. Which tasks did it take a shortcut on?
  2. Switch Permissions to Branch only. What did taking away main and production stop, and what did it miss?
  3. Choose Branch + approval. How much of a person's time does it cost, and do routine tasks wait?
  4. Set Permissions back to Full access and turn on Give it Initech's rules. Does a written rule hold?
  5. Keep the rules on and choose Branch + approval. What reaches the approver now?
  6. Turn the rules off, set Full access and switch to Fast. Is the smaller model more careful?
askyesnorunTasks10 in Initech's repoCoding agentstub tools, nothing runsmistral-smallPermission gaterun, block or ask?ApproveroffRepo and prodthe call would runScorecardrisky vs. done
Press Run to watch it run

Results

Run it to see what the agent reached for on each task.

Behind the scenes

Run it and every task, each risky tool call the agent made, and what the gate did with it will show up here.