Agents release notes

You can test different model configurations directly within evaluation sets. Configure multiple model and temperature combinations for an evaluation set, run them in parallel, and compare accuracy, latency, and behavior across identical test cases. This update makes it easier to identify the best-performing model setup for your agent before deploying changes to production.

For details, refer to Choosing the best model for your agent.

We are introducing feedback on agent traces, a new capability that allows you to annotate and review specific parts of an agent’s runtime. You can now provide feedback on individual spans, including tool calls, guardrail checks, and model outputs, making it easier to diagnose issues, highlight correct behaviors, and guide iterative improvements.

You can now define a maximum number of iterations for agents directly in agent settings. This limit helps you control cost, manage execution time, and prevent infinite loop behavior.

The Max. iterations setting represents how many reasoning loops an agent can perform before it must either complete the goal or stop. You can set the value anywhere between 1 and 100, allowing you to tailor it to simple or highly complex use cases. Once the agent reaches the configured limit, it must finalize the task or halt execution.

This setting is available only for autonomous agents, not conversational agents.