A Subagent Opens A New Context

Harness Series Article 6, about subagent.

A subagent is a tool call that opens a new work context for the coding agent.

What happens when a subagent is launched?

Short answer: Tool call outside, runtime inside.

First, the Tool Call

Why is launching a subagent essentially a tool call?

For example, the user says: “Use a verifier subagent to audit webhook retries.”

If Harness supports it, you can directly use the slash command /delegate to make a spawn_agent tool call.

After the tool finishes, the parent gets a child handle.

Now, the Runtime

In Harness article 3, we mentioned that every tool call changes the coding agent’s world state.

So what happens after this tool call that launches a subagent? Let’s look at the runtime inside.

The tool call is the entry point. Going deeper, Harness opens a new work context.

What is the relationship between a subagent’s context and the coding agent’s context?

Session, Context, Subagent

These three terms are easy to mix up.

• Session is the runtime container: thread, transcript, tools, permissions, resources, status, artifacts.

• Context is the projection visible to a given model call: instructions, skills, AGENTS.md, recent turns, summaries, tool results, file state.

• Subagent is a child session opened under the parent session. It can inherit resources and also receives a selected context slice.

Tool Call Outside, Runtime Inside

• The term “subagent” can be a bit misleading—it sounds like a smaller agent.

• In Harness, it’s more like a managed child runtime. A subagent can use many of the same Harness resources the parent can: tools, skills, AGENTS.md, MCP servers, cwd, sandbox, permissions.

• But shared resources do not mean a shared transcript. The child has its own work context. The parent decides how much context to project into it.

Fresh Agent, Forked Agent, or Partial Fork?

What context the parent projects to the subagent depends on the subagent type. There are three common subagent patterns:

• Fresh child – It needs to receive: goal, relevant files, what has been tried, exact output, and depth of answer.

• Forked child – It has already inherited the surrounding context, so the prompt should directly give the next directive.

• Partial fork – This is the most practical intermediate option. It gives the child enough local memory to work while preventing the parent history from becoming inherited noise.

Here are some examples to think about: what kind of subagent suits which task?

• Parallelism Agent: one subagent checks database migration, one looks at frontend state, one runs verification.

• Role Specialization: Explorer, Planner, Verifier, Worker, Reviewer…

• Background Work: e.g., multiple subagents handle a large refactor or long tests.

Subagent Workflow

A practical subagent workflow:

Organizing multiple subagents to accomplish work is actually very challenging. The clearer the definition of a subagent’s role, context, and tools, the better.

Finally:

More agents do not guarantee better work; more agents create more runtime state.

Harness must know who is working, what it knows, what it changed, when it finishes, and how the result becomes evidence.

A subagent expands the world for the coding agent, and Harness is about whether that new world is better or more chaotic.