A Native Harness for Agents That Need to Finish the Job
Most agents are built to answer a turn. Many useful jobs are not. Today we added the FabrCore Agent Harness: a runtime composition that gives any host agent a model-managed todo list, an iteration loop, durable progress, and delegation without asking developers to build a second orchestration system.
Composition, Not a New Base Class
The Harness wraps the agent you already have. Create one directly with CreateFabrCoreHarnessAgent, or compose an existing Microsoft Agent Framework agent with AsFabrCoreHarnessAgent. Your tools, middleware, model choice, instructions, telemetry, and FabrCore identity remain intact.
That was a non-negotiable design constraint. Developers should not have to choose between a conversational agent and a task-capable agent at the inheritance boundary. The Harness is behavior added to a host, which means it can be enabled where the workload needs persistence and omitted where a single model turn is enough.
The Todo List Is Model-Managed
The Harness exposes a todo list as part of the agent's working contract. The model can add, update, complete, and reorder items while it learns more about the objective. Each iteration sees the current plan and the evidence produced so far, then chooses the next useful action.
This is more than asking the model to "think step by step." Todo state is structured, observable, and persisted. A host can report progress without scraping prose, and a reactivated grain can restore the work instead of hoping the prompt recreates the same plan.
A Loop with an Exit Contract
Autonomy without an exit contract is a runaway bill. The Harness loop has explicit completion semantics and bounded iterations. It stops when the todo list is complete, the model produces the final result, cancellation arrives, or configured safety limits are exhausted.
The important move is that these controls live beside FabrCore's model configuration and context budgets. The same runtime that knows the agent's history and token use also knows whether another iteration is safe. Tomorrow we will show how the new compaction ladder makes that relationship concrete.
The Agent Harness guide includes the minimal agent, modes, configuration arguments, and production checklist.