A new technique repairs AI agent instructions by reading what the agent already did, not by paying it to try again.
When a system prompt is vague, the same input can send an agent down different paths and produce inconsistent results. The usual fix, prompt optimization, generates several candidate prompts and runs the agent on each to score them - doubling the cost, since you pay for rollouts once to generate and again to rank. A new method called Semantic Feature Analysis (SFA) skips that ranking step: it clusters outputs from traces the agent already produced, extracts the features that best explain good versus bad outcomes, and writes the winners back into the prompt as corrective instructions. Tested against five existing optimizers - GEPA, MIPROv2, SIMBA, BootstrapFewShot, and InferRules - across four benchmarks at matched dollar budgets, SFA beat the unmodified agent every time, including on GAIA, where rivals couldn't afford to test a single candidate prompt.
That's the real bottleneck as agents get wired into expensive tool calls and long chains: not creativity, but the electric bill for testing ideas. A method that improves agents from trace data alone, without spending on extra rollouts, targets the constraint teams actually hit in production rather than a benchmark-friendly one.
Whether SFA holds up outside these four benchmarks, and how well its decision tree scales to messier real-world traces, is the question the paper leaves open.