agent-auto-improvement

Pass

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: SAFE
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill references documentation hosted at 'docs.braintrust.dev', which is the official domain for the Braintrust platform described in the instructions.
  • [COMMAND_EXECUTION]: The instructions involve the use of the 'bt' command-line interface. This is the standard tool for the Braintrust platform and is used here for legitimate operations such as querying project logs, running datasets, and performing evaluations.
  • [CREDENTIALS_UNSAFE]: The skill correctly advises users to manage their authentication via the 'BRAINTRUST_API_KEY' environment variable in CI environments or via the 'bt setup' command, adhering to standard security practices for secret management.
  • [INDIRECT_PROMPT_INJECTION]: The skill inherently processes external data (production traces).
  • Ingestion points: Production data is ingested via project_logs and dataset SQL table functions in SKILL.md.
  • Boundary markers: The skill explicitly recommends using binary, deterministic code-based scorers to evaluate hard failures rather than relying solely on LLM-as-a-judge, which serves as a security boundary.
  • Capability inventory: The skill executes agent logic through the task function in the Eval configuration to compare outputs against baselines.
  • Sanitization: The workflow includes advice on normalizing message histories (e.g., dropping orphaned tool blocks) before replaying them to the model, which helps mitigate structural data issues.
  • [DYNAMIC_EXECUTION]: The skill uses tsx or vite-node to run TypeScript pipelines and evaluation files. This is the expected execution model for the Braintrust SDK and is used to process the user's local code.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 3, 2026, 05:47 AM