Sub-agent
A secondary AI agent that a main agent starts to handle part of a task, such as research or a code change, often with its own context and tools. It is part of the harness around the model.
A sub-agent is an agent launched by another agent to do a defined piece of work. The main agent hands over a narrow task, the sub-agent works with its own context and tools, and it returns a result. This keeps the main context focused and allows parallel work.
Sub-agents are one of the components of the harness around a model, along with context management, tool use and the execution environment. Providers increasingly treat these as parts of the system that evolve alongside the model itself.
Implications
Because sub-agents multiply tool calls, actions and cost, they also multiply risk. Each has its own permissions and failure modes, and an error can compound as it passes between agents. For evaluation, the configuration matters as much as the model: how many sub-agents, with which tools and limits (runtime address). For governance, define what each is allowed to do, since capability is not authority. See agentic AI.
Read more in The benchmark needs a runtime address.
Related terms
Harness
The software around a model that manages context, tool use, sub-agents and the execution environment. It shapes real-world capability, so it must be evaluated with the model, not ignored.
Agentic AI
AI systems that pursue a goal by planning, using tools and taking multi-step actions with some autonomy, rather than only answering a single prompt.
Tool call
When an AI model asks to run an external function, such as searching, reading a file or updating a record. A failed tool call can be a very different failure from a wrong answer.
Used in these essays
Your AI agent estate is becoming a capital allocation problem
Once agents multiply across workflows, the management problem changes: which agents deserve more capacity, which should be merged or constrained, and which should be retired?
The benchmark needs a runtime address
AI systems are increasingly dynamic at runtime. Enterprise evaluation should qualify the serving route, harness, tools and fallback conditions, not just the model name.