The Agent Skills Directory

[SAFE]: The skill provides comprehensive architectural guidance for observability and evaluation loops. It emphasizes security-conscious patterns such as the redaction of sensitive identifiers and the use of deterministic verification steps to prevent agents from gaming evaluation suites.
[INDIRECT_PROMPT_INJECTION]: The skill identifies architectural surfaces where untrusted data, such as file diffs and tool outputs, enters the evaluation context. It provides detailed remediation guidance, including the use of structured rubrics and data redaction to mitigate injection risks.
Ingestion points: Untrusted data enters the agent context via session traces, replayable transcripts, and file diffs as described in SKILL.md and references/trace-and-telemetry-model.md.
Boundary markers: The skill suggests using structured judge prompts and scoring rubrics (references/evals-regression-and-cost-ops.md) to isolate and grade outputs.
Capability inventory: The skill references external execution runtimes and includes examples of shell command execution within agent tasks.
Sanitization: The documentation explicitly requires hashing or redacting user-identifying data and proprietary source code before telemetry export (SKILL.md, references/trace-and-telemetry-model.md).
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill provides specific advice on avoiding data-residency liabilities by redacting proprietary source code and secrets from centralized trace stores. It also details the use of W3C trace propagation to maintain observability without leaking PII into metric dimensions.
[DYNAMIC_EXECUTION]: The skill discusses 'Harness Self-Evolution' as an advanced technique where an agent proposes edits to its own harness components. To mitigate risk, the documentation prescribes strict controls, including versioned component files, pre-declared effect predictions, and mandatory human-in-the-loop release gates identical to those for human-authored changes.

ai-coding-agents-observability-evals