Evals from code: Test agent changes the way the agent actually runs

AgentControl’s Evals from code feature enables teams to evaluate AI agents through their own Python or Node.js application code, allowing tests to use the same tools, retrieval processes, providers, configurations, and handlers as production deployments. This approach addresses limitations of prompt-and-model-only evaluations, which can miss behavioral regressions such as an agent failing to call a required order-lookup tool despite producing a plausible answer. Evaluations can combine model-based AgentControl judges for qualities such as accuracy and tone with custom code scorers for deterministic checks, including tool-trajectory reviews that assess whether tools were called appropriately. Runs can use the exact config variation intended for release, report pass or fail for CI gating, and display row-level verdicts, reasoning, cost, and latency in the Evaluations page. The feature complements Playground evaluations, which are intended for comparing prompts and models independently, while evals from code test the full production-like agent workflow. It requires version 1.0 or later of the LaunchDarkly AI SDK and currently supports Python 3.12+ and Node.js 22+, with setup available through AgentControl and integration into CI systems.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.