Agent harness glossary

Short definitions for agent harness engineering terms: agent harness, harness engineering, agent observability, agent evals, guardrails, and trajectory evaluation.

• Agent harness

  An agent harness is everything in an AI agent except the model: the tools, context, prompts, memory, hooks, guardrails, and feedback loops that turn a model into a working agent.

• Harness engineering

  Harness engineering (or agent harness engineering) is the discipline of designing, measuring, and improving everything around the model in an AI agent so the agent is reliable in production.

• Agent observability

  Agent observability is the practice of tracing and monitoring the full agent harness: LLM calls, tool calls, MCP requests, retrieval, multi-step trajectories, cost, latency, and outcomes.

• Agent evals

  Agent evals are systematic checks that score agent quality, safety, and cost on real traces or offline datasets—using LLM-as-a-judge, programmatic rules, or human review.

• Guardrails

  LLM guardrails are runtime controls that detect or block unsafe prompts and responses—such as prompt injection, sensitive topics, and topic restriction—before they reach users or tools.

• Trajectory evaluation

  Trajectory evaluation (or tool-call evaluation) scores the sequence of steps an agent took—tools chosen, arguments, ordering, and intermediate decisions—not only the final response.