Definition
Plain language
A wrapper around an AI agent that catches common mistakes before they reach the outside world.
As stated in the literature
A four-layer runtime harness (action realization, environment contract, trajectory regulation, procedural skill) evolved via Codex from a small model's failure traces and shown to transfer across model families and scales.
Why it matters: It shows that wrapping a weak model in good runtime checks can recover much of the reliability of a stronger one, and that those checks transfer across models.
For example, before an agent's 'rm -rf /tmp/build' reaches the shell, the harness checks it against an action-realization rule and rewrites or rejects the command.
Heard on the show
“But now — take that base Qwen, the one xLAM started from, and wrap it in Life-Harness.”Episode 071 — When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface