Subagent
A model instance spawned by another to handle a scoped task with its own context.
A subagent is a model instance created by another agent to perform a delimited task and return a result. The motivation is usually context economy: a search that would flood the parent's Context Window with file contents can be delegated, with only the conclusion coming back.
Delegation has costs that are easy to underestimate.
The parent cannot see the work, only the report. A subagent that claims a build is green, a change is pushed, or a check has passed is making a claim, not providing evidence — verify independently. See Ground Truth.
Scope leaks. A subagent inheriting the parent's full context frequently takes on more than it was assigned, because it can see the whole problem.
Nesting is usually capped, so a design that assumes a delegate can itself delegate may silently degrade — commonly into self-review, which is exactly the perspective Adversarial Review exists to avoid.
Isolation is not automatic. Unless given its own working copy, a subagent shares the parent's directory, and destructive commands there destroy in-flight work.
The instruction that most reliably improves subagent output is a statement of what "done" means and what evidence must accompany it.
See also5
Hand-picked in the note itself — the neighbours worth reading next.
Multi-Agent Orchestration
Coordinating several model instances on one body of work, each with its own context.
Agents & Language Models10 connections
Adversarial Review
Reviewing work with an explicit mandate to refute it, from a perspective that did not produce it.
Agents & Language Models9 connections
Git Worktree
An additional working directory attached to one repository, each on its own branch.
Version Control & Delivery10 connections
Token Budget
The finite allowance of model usage a task may consume, and the design decisions it forces.
Agents & Language Models11 connections
Fan-Out and Fan-In
Splitting work across parallel workers and recombining their results.
Agents & Language Models9 connections
Related2
Nearby in the graph rather than deliberately chosen. Looser, sometimes surprising.
Linked from8
Notes elsewhere in the wiki that reach for this one.
- Adversarial ReviewAgents & Language Models
Reviewing work with an explicit mandate to refute it, from a perspective that did not produce it.
- Context WindowAgents & Language Models
The bounded span of tokens a model can attend to in a single request.
- Fan-Out and Fan-InAgents & Language Models
Splitting work across parallel workers and recombining their results.
- Git WorktreeVersion Control & Delivery
An additional working directory attached to one repository, each on its own branch.
- LLM-as-JudgeAgents & Language Models
Using a language model to score another model's output against a rubric.
- Multi-Agent OrchestrationAgents & Language Models
Coordinating several model instances on one body of work, each with its own context.
- Resource StarvationTesting & Verification
Failures caused by competition for a finite shared resource rather than by any defect.
- Token BudgetAgents & Language Models
The finite allowance of model usage a task may consume, and the design decisions it forces.