evaluate-and-harden-agentlisted

# Skill: evaluate-and-harden-agent > **Invoked by:** `agent-implementation-engineer` (primary). Also consulted by `agentic-systems-architect` when a go/no-go needs an evidence bar defined before build. > > **When to invoke:** "how do I know this agent actually works?"; "set up evals for the agent"; "the agent loops forever / burns tokens"; "the agent did something irreversible it shouldn't have"; "it works in the demo but fails in the tail". > > **Output:** an offline agent-eval harness (task-completion + trajectory + tool-use, over a fixed task set) and a hardening plan (caps, timeouts/retries, stop conditions, human-in-the-loop, tracing) — with cost and latency reported next to quality. ## Procedure 1. **Build a fixed task set before scoring anything.** Collect representative tasks with known-good outcomes (start with real or realistic inputs, include the hard/tail cases, not just the happy path). This set is the regression harness — freeze it and grow it as failures surface. 2. **Score three dimensions, not one.** - **Task-completion:** did the agent achieve the goal? (the outcome, judged against the known-good result — exact-match, rubric, or LLM-judge as fits.) - **Trajectory:** did it take a *sane path*, or flail and get lucky? Check step count, redundant/looping calls, and whether the plan was coherent — a right answer via a 40-step random walk is a latent failure. - **Tool-use correctness:** right tools, right arguments, recovered from errors? Wrong-tool a