Design agent evaluations around reviewed tasks, controlled environments and explicit verifiers.
Editorial note checked 2026-10-29 against this instruction snapshot. Supporting scripts and runtime behavior remain outside this editorial review.
The workflow inspects a repository and optional traces, develops task specifications and builds reusable evaluation artifacts. It distinguishes task generation from running and auditing an evaluation.
Teams measuring a concrete agent capability or regression.
Synthetic tasks can favor the implementation being tested. Evaluation dependencies and external services need review, and a verifier can reward behavior unrelated to user success.
Unknown; inspect source requirements and supporting scripts.
Inspect the pinned source snapshot and its supporting files first. The command selects this skill through the open-source skills CLI. Node.js and agent-specific setup may be required. Installation can download code. Catalog ingestion never executes skill instructions or install scripts.
npx skills add https://github.com/langchain-ai/langchain-skills/tree/126c384d8552dcb476b226c244172186cc3f4328/config/skills/eval-engineering --skill eval-engineeringReview pinned skill folder ↗ · CLI documentation ↗
After source review, use npx skills check and npx skills update to inspect and apply upstream updates. Re-check changed instructions before running them.
Static review only. This is not a security audit or a guarantee of safety.
SKILL.md text only; scripts, dependencies and runtime behavior not audited.
No configured text patterns matched. Unmatched or concealed behavior can still exist.
Allowed tools declaration: Unknown. Runtime permissions and network behavior: Unknown.
86/100 · Metadata coverage 100% · Model 1.0
Scores summarize observable discovery signals. Popularity and repository activity apply to the whole repository. Unknown runtime safety earns no certification.
Define the task and acceptance criteria first, review the environment and construct a small auditable evaluation set.
Run known good and bad behaviors, inspect verifier blind spots and retain task versions and evidence before comparing model or system scores.
Related by category; compare scope and dependencies before choosing an alternative.
Explore dispatching parallel agents as a productivity & planning workflow from obra/superpowers. Compare its source, maintenance and review signals before choosing it for your agent.
Explore executing plans as a productivity & planning workflow from obra/superpowers. Compare its source, maintenance and review signals before choosing it for your agent.
Explore finishing a development branch as a productivity & planning workflow from obra/superpowers. Compare its source, maintenance and review signals before choosing it for your agent.
Configure a Deep Agents application around its harness and supported extension points.
Choose Deep Agents storage backends according to persistence and filesystem requirements.
Coordinate Deep Agents tasks with explicit delegation, planning and approval boundaries.