Fine-grain vs coarse-grain agents
Fine-grain agents are strict, typed, and predictable. Coarse-grain agents plan and execute broad goals with more autonomy.
Fine-grain micro-agents
Definition: very small, single task, strict schema, minimal discretion. Examples: “extract invoice total”, “classify intent”, “redact PII”.
Pros
Highly testable + predictable; cheap + fast; easy to swap/upgrade; easy to secure (limited tool access).
Cons
More orchestration complexity; more hops can add latency if poorly designed.
Use when
You need structured outputs, compliance matters, workflows repeat, and reliability at scale is required.
Coarse-grain agents
Definition: bigger scope, can plan multi-step work, more autonomy. Example: “handle customer onboarding end-to-end”.
Pros
Fewer handoffs; good for ambiguous problems; faster to prototype.
Cons
Harder to test and govern; more prompt drift; debugging becomes archaeology.
Use when
Domain is ambiguous, early prototype stage, and a single owner agent makes sense behind policies.
Recommended hybrid model
Coarse-grain for planning
Use coarse-grain agents for intent understanding + planning.
Fine-grain for typed core
Use fine-grain agents for extraction, validation, compliance, and final formatting.
Workflow glue
Coarse-grain agents orchestrate. Fine-grain agents are the typed core.