ADK-built agents
Agents written with Google's Agent Development Kit, often with several cooperating agents and custom tools, deployed to managed runtimes such as Agent Runtime.
SERVICES / CLOUD PLATFORMS
Evaluate agents built on Google Cloud.
Independent testing and evaluation for Gemini-based agents built with Gemini Enterprise Agent Platform and the Agent Development Kit, assessing both the path an agent takes and the outcome it delivers.
Common Google Cloud agent patterns
Gemini Enterprise Agent Platform
Agents and conversational experiences configured on Gemini Enterprise Agent Platform (formerly Vertex AI), connected to enterprise search, data stores and APIs.
Gemini in custom applications
Applications calling Gemini models directly, with retrieval, function calling and long-context document work.
Quality issues we look for
Right answer, wrong path
The agent reaches a plausible result through unnecessary or incorrect tool steps, which breaks once conditions change.
Lost context in long sessions
Details provided earlier in a conversation or document are missed, contradicted or overwritten later.
Over-reliance on model knowledge
Answers drawn from general model knowledge instead of your approved sources, creating confident but unsupported claims.
What we assess
Trajectory evaluation
Compare the agent's actual sequence of tool calls with acceptable reference paths, allowing valid alternatives where your workflow permits them.
Final response and outcome
Score the final answer with calibrated rubrics and verify the resulting state in the systems the agent touched.
Model and version comparisons
Compare Gemini model versions and configurations on the same reviewed cases, with quality, latency and cost side by side.
Working with native Google Cloud evaluation
ADK and Gen AI evals
Google provides evaluation support for agent trajectories and responses through ADK and Gen AI evals on Agent Platform. We supply the reviewed datasets, reference paths and rubrics that make those scores meaningful.
Expert calibration
Automated metrics are compared with domain reviewer labels before release thresholds are set.
Reproducible runs
Agent versions, model settings and data snapshots are recorded with each run so results can be repeated and compared.
Plan the engagement
Your inputs
A test project or environment, agreed access, deployment and version details, business rules, example tasks and a domain reviewer.
Your deliverables
An agreed benchmark, prioritized findings, reusable evaluation assets and recommendations for the next iteration.
Support after handover
We can retest fixes, extend coverage to new Google Cloud workflows or move to recurring Managed Evaluations as your agent changes. Platform usage and licenses stay in your account.
Make your next decision with evidence.
Start with one workflow, clear success criteria, and an evaluation plan your team can use.