Agent, task, and executor
Public Preview

Understand the agent harness contract

The agent harness coordinates model calls, tools, and delegated executors between one task submission and one terminal outcome.

The agent harness coordinates model calls, tools, and delegated work so your application submits one task and receives one terminal outcome within a set deadline.

  • Defines agent, task, and executor as the three objects that structure the work.
  • The harness manages model routing, built-in tools, registered functions, and child executor delegation.
  • The task contract stays stable regardless of the executor tree shape the harness chooses.
  • Points to Executors and deadlines and Context continuity as the next guides to choose.

Summaries were generated by AI. Generative AI is experimental.

Your application submits one task and receives one terminal outcome. Between those points, the agent harness coordinates models, tools, and delegated work within the deadline you set. Stored history and result handoffs keep context available as the work progresses.

Your process handles calls to the tools it registered. Duale AI coordinates the rest, so your application does not call a model provider or join delegated results itself.

Agent, task, and executor

These objects describe different parts of the work.

ObjectMeaning
AgentThe identity, tools, and access your application uses to submit work
TaskOne submission with an instruction, an absolute deadline, and one terminal outcome
ExecutorOne bounded unit of agent work. The root carries the conversation; children handle delegated work
  • Object
    Agent
    Meaning
    The identity, tools, and access your application uses to submit work
  • Object
    Task
    Meaning
    One submission with an instruction, an absolute deadline, and one terminal outcome
  • Object
    Executor
    Meaning
    One bounded unit of agent work. The root carries the conversation; children handle delegated work

Agents and access defines agent identity and access. SDK concepts defines the task and result contract.

Your application submits a task for an agent. The harness creates the executors needed to complete the task.

What the harness coordinates

Duale AI coordinates the work needed to produce the terminal outcome:

  • model calls through eligible routes, guided by the task’s routing preferences;
  • built-in document, web, and computation tools;
  • calls to functions your process registered with @tool;
  • delegation to child executors and the return of their results;
  • stored task history, the active model context, and targeted recall.
The managed loop from one task to one terminal outcomeYour application submits one task to a root executor. The executor can use built-in tools, call a tool registered by your process, or delegate focused work to a child executor. Child results return to the root, which produces the terminal outcome for your application.

The available sources depend on the agent’s access and tools. Authoring tools defines the functions your process supplies. Libraries and Web search explain the document and web sources.

One stable application contract

The harness can use different bounded executor trees for similar work without changing the task contract. Your application provides the instruction, response model, deadline, and routing preferences, then receives one terminal outcome.

SDK concepts defines the task and terminal-outcome contract. Build your integration around that contract and let the harness choose the executor-tree shape.

Choose an Agent harness guide

Agent, task, and executorWhat the harness coordinatesOne stable application contractChoose an Agent harness guide
Last updatedAugust 31, 2026