Definition
Plain language
A research coding agent that edits its own instructions to get better, with a panel of AI reviewers voting on each change.
As stated in the literature
Self-Improving Coding Agent: a loop that reads its own transcripts, proposes edits to its coding agent's natural-language prompt, and gates them behind a committee of LLM reviewers; compromised within one to two generations by the poisoned benchmark.
Also called: Self-Improving Coding Agent
Why it matters: It is a concrete instance of software that edits its own operating instructions, which means any flaw it picks up can be carried forward into every later version of itself.
For example, after failing a batch of tasks, SICA reads back its own transcripts, rewrites a line of its instructions to say it should try a different approach, and only keeps the edit if its panel of reviewers approves.
Heard on the show
“SICA, the Self-Improving Coding Agent, edits the natural-language prompt of its own coding agent, and runs every proposed change past a committee of three language-model reviewers.”Episode 268 — A Rigged Benchmark Taught a Self-Improving Agent to Always Disable SSL