looping

Agent looping

Agent looping re-invokes an agent until a completion condition is satisfied. Use it for iterative refinement, todo completion, waiting for background tasks, or evaluating whether an answer meets explicit criteria.

Always bound autonomous loops. A completion condition can fail, a model can stall, and an evaluator can be probabilistic.

[!IMPORTANT] Agent looping is experimental.

Set up looping manually

Use the direct composition API when you want looping without the other Harness Agent defaults.

::: zone pivot="programming-language-csharp"

Import the loop types and wrap any AIAgent with LoopAgent. Its default maximum is 10 agent invocations:

using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;

AIAgent baseAgent = chatClient.AsAIAgent();
AIAgent agent = new LoopAgent(
    baseAgent,
    new CompletionMarkerLoopEvaluator("DONE"),
    new LoopAgentOptions { MaxIterations = 5 });

::: zone-end

::: zone pivot="programming-language-python"

Import AgentLoopMiddleware and add it to a regular Agent. The default maximum is 10 agent runs:

from agent_framework import Agent, AgentLoopMiddleware


def needs_more_work(*, last_result, **kwargs):
    return "DONE" not in last_result.text


agent = Agent(
    client=client,
    middleware=[
        AgentLoopMiddleware(
            needs_more_work,
            max_iterations=5,
        )
    ],
)