Microsoft's Agent Harness ends an autonomous run when the model types RESEARCH_COMPLETE

Call `AsHarnessAgent()` on an `IChatClient` and you get an `AIAgent` back ^[3]. It is the same type as the flat agent that takes one message, possibly calls tools, and answers ^[12]; the extra behaviour is switched on by configuration on that object ^[4]. According to Isaac Ojeda's dev.to walkthrough, the Harness composes pieces already in the framework instead of defining a parallel runtime ^[2][20].

The chat pipeline supplies function calling, message injection and history persistence ^[9]. Context providers carry instructions, memory, todos and operation modes ^[10]. Tool approval and observability are middlewares, and so is the bounded loop, which the post lists as optional ^[11].

The evaluator does the deciding. `CompletionMarkerLoopEvaluator` reads the model's last response and looks for the marker string exactly; while the marker is absent, the loop invokes the agent again ^[8].

A model that finishes its report without typing the marker keeps going to the cap, so the run costs the full `MaxIterations` in model invocations ^[18]. A model that puts the marker in its last response for any other reason stops there ^[21]. The only ceiling the post describes is that iteration count; there is no token or spend budget in the options it shows ^[19].

`HarnessInstructions` and `ChatOptions.Instructions` are separate fields, one for how to behave as an agent, including autonomy and tool discipline, the other for the domain role ^[5]. The harness block is always prepended to the domain block ^[6]. The ordering is fixed by the framework, so a domain instruction that contradicts the agent-behaviour text arrives second in the prompt.

Whether this absorbs the scaffolding in your codebase depends on the shape of your stop condition. If the agent can finish on a marker the model emits, or on a custom evaluator over the assistant's last message, the Harness loop replaces your `while` and your safety limit becomes an iteration count ^[13]. If your ceiling is denominated in tokens or dollars, or your completion test needs to read tool results, that code is still yours to maintain.

This is one practitioner walkthrough, not the framework documentation. It is honest about which decisions it hands over: Ojeda never tells the sample how many searches to run or in what order, and the agent decides iteration by iteration ^[16]. The full sample is published on GitHub as TopicResearchAgent ^[17].