Google Cloud shares four evaluation engineering lessons from building agent plugins

Building reliable AI agents requires rigorous evaluation engineering rather than just construction, according to Google Cloud. Using the newly released Google Cloud Developer plugin as a testbed, the team identified four critical lessons for measuring agent behavior in real-world workflows.

Grade Execution Trajectories, Not Summaries

Initial evaluations using an LLM-as-judge on final text replies proved flawed because the grader could not see actual actions. In an 11-prompt evaluation suite, the automated grader marked three prompts as failing, but raw tool-call logs revealed the agent had successfully executed required commands, such as gcloud help run services list, without mentioning them in the final summary. The takeaway is to score the raw execution trajectory to verify what commands were actually run and validated.

Preserve Turn Structure in Telemetry

Extracting agent telemetry into a flattened, sequential list obscures concurrent tool dispatch, a key cause of behavioral failure. Modern agent harnesses often batch multiple tool invocations into a single assistant turn. In one case, a skill-loading call and two documentation searches were dispatched simultaneously, causing the skill's instructions to arrive in the context window after the searches had already executed. Engineers must preserve turn structure in telemetry extractors to identify when parallel dispatch issues tool calls before injected guidance enters the context.

Run Live and Plan-Only Tests Side-by-Side

Testing in a single environment condition is insufficient. The team recommends running two conditions simultaneously: a live run with sandboxed credentials to test actual execution and safety restraint, and a plan-only run with stripped credentials to expose the model's raw reasoning and routing intent. Roughly half of the most important findings came from the delta between these two conditions. For example, a live run blocked by a permission error masked a major cognitive flaw that a plan-only run exposed in seconds, revealing that the agent ignored organization policies and governance.

Use Golden Sets as Durable Baselines

As frontier model architectures and agent harnesses evolve, evaluations should function as continuous integration testing for the non-deterministic stack. Instead of guessing from final text, teams should capture the whole trajectory—thoughts, commands, and outcomes—and score them against golden sets curated by human experts. These baselines serve as a safety net against model churn, ensuring the composite system of harness, model, and tools continues to perform as expected.