Nobody told it to
A model in an internal evaluation did something nobody asked it to do, and the interesting part is what that does and does not prove.
Tags
9 pages
Agents
House rules
Running three different coding agents against one repository, and the house rules that stopped them undoing each other's work.
The afternoon the agent earned its keep
Adding stream-copy remux to a WebAssembly FFmpeg build, where writing Matroska hung forever and the identical MP4 path did not.
The off-switch was never a button
The kill-switch answer to AI autonomy does not survive contact with how these systems actually run. Governance is not a button.
The agent said SUCCESS. The linter disagreed.
An AI repair agent declared success the moment the code compiled. The linter disagreed, and the gap between those two is the whole story.
The interpreter we forgot to sandbox
Hundreds of malicious package versions across three ecosystems, and the uncomfortable fact that installing one runs arbitrary code.
The goblin that wouldn't stay dead
Building an AI dungeon master to test a Go chat client, and the bug that kept resurrecting a goblin the player had definitively killed.
An AI agent that has to make the build pass
Generate-and-hope does not work for codegen. Handing the draft to a repair agent with a fixed toolset that has to make build, test and lint pass.
Letting the AI call your Go functions
Tool calling and the reason-act-observe loop explained, and how to define a tool in Go so the schema comes from the type rather than by hand.
