AWS gives its system prompt optimizer a shell over a directory of scored traces

The context window forces the filesystem. AWS says agent traces run long enough that even a few dozen can exceed a model's context window, so passing every trace into the reflector's prompt is impractical ^[7]. The two obvious workarounds, truncation and pre-summarization, both decide in advance which parts of a trace matter. AWS skipped both and made the complete trace corpus available to the reflector through a filesystem ^[8].

AWS describes a minimal workflow. An evaluator scores a collection of traces and writes them to a directory; the reflector is given a shell tool, pointed at that directory, and instructed to propose configuration updates by examining patterns of success and failure ^[9]. From there it lists files, searches with grep, reads traces with cat and compares outputs with diff ^[10].

The reflector determines which evidence matters, what comparisons to make, and how to turn those findings into configuration changes; no fixed signal-extraction or trace-summarization pipeline is imposed ^[11]. I would make the same call for this failure mode. A platform-built summarizer has to guess which fields of a trace carry the failure, and it guesses once, for every customer. It costs reproducibility. A reflector that picks its own greps can pick different ones next time, and the guardrails screen the proposed edits while the search path that produced them goes unchecked ^[12].

Epochs multiply that. Each optimization epoch repeats the same cycle: score the traces, reflect, accept the edits that pass the guardrails ^[15]. Run three and the evaluator has scored the corpus three times and a reflector has read through it three times ^[1]. AWS puts the extra epochs as trading optimization time for further quality gains ^[15].

The customer owns the reward signal. The optimizer reads traces recorded in AgentCore Observability together with a reward signal to produce the improved prompt ^[3], and the revised prompt will chase whatever that scorer rewards. Observability supplies the behaviour and evaluations supply the quality signal ^[2].

The published benchmark comparisons are claims about the benchmark's agents. For the ranking to hold on yours, your traces would have to fail the way those traces fail, and your evaluator would have to score the way that benchmark's scorer scores.

For the market trends agent, the recommendation compares the existing system prompt with a revised version and explains the trace patterns that motivated the changes, according to AWS ^[4]. The interface includes an explanation of the proposed changes to review before deciding whether to test the recommendation ^[5]. AWS says recommendations should be reviewed and tested before use, and that the guardrails screen candidate updates before they can be promoted ^[13].