The AI Agent Evaluation Playbook
How to evaluate agentic systems where success is multi-step, non-deterministic and hard to grade. A working framework from production deployments.
Why agent evals are harder than LLM evals
A single-turn LLM eval scores an answer against a reference. An agent eval has to score a trajectory — a sequence of tool calls, intermediate reasoning and final result — where many valid paths exist. Static ground-truth doesn't work when the agent could reasonably take different but equally correct routes to the same outcome.
The four eval layers we ship with every agent
Layer 1: task completion — did the agent achieve the goal? Layer 2: efficiency — how many tool calls and tokens did it burn? Layer 3: correctness of intermediate steps — did it call the right tool with the right args? Layer 4: safety — did it stay inside its allowlist, respect cost caps and escalate when it should have? All four run on every eval batch, not just the final answer.
Golden trajectories vs. LLM-as-judge
For narrow tasks, curate 50–200 golden trajectories with expert-labelled tool calls. For broad tasks, use LLM-as-judge with a rubric — a stronger model scores whether the agent's plan and result are sensible. Combine both: golden for regression on known-hard cases, LLM-as-judge for coverage on the long tail.
Continuous evaluation in production
Sample 1–5% of production traces daily and score them. Alert on regressions per task type, tool and model version. Route ambiguous cases to human review — those become the next round of golden trajectories. Evals aren't a launch gate; they're a rolling QA function.
What good looks like
For a research agent, we hold p95 task completion above 85%, cost per completed task under a target, and safety violations near zero. Regressions of >2 points on any of the four layers gate the deploy. Every prompt or model change goes through this loop before it touches production traffic.
Key takeaways
- Evaluate trajectories, not just final answers
- Score four layers: completion, efficiency, step correctness, safety
- Combine golden trajectories with LLM-as-judge for coverage
- Sample 1–5% of production traces daily; regressions gate deploys
- Ambiguous cases feed the next round of golden data
T7 Research
Enterprise AI Research Group
T7 Research is the research arm of T7 Solution, focused on benchmarking LLMs, evaluating RAG patterns, and compiling implementation playbooks for enterprise technology leaders.