Orchestration Patterns — How to Compose Model Calls

"Ask the LLM and get an answer" — fine for simple questions. But "fix this repo's bug" doesn't work in one shot. Read files, analyze the problem, create a patch, run tests, fix again if they fail. Designing this entire flow is orchestration.

Key Patterns

Chaining — serial structure where the first call's output becomes the second call's input. Simplest but effective. Example: code generation → code review → revision. Each step can use different prompts.

Routing — analyze input and send it to the appropriate processing path. "This question is code-related, send to coding agent." "This is docs-related, send to search agent." A classifier model acts as router.

Parallelization — run independent tasks simultaneously. Analyzing 10 files is faster in parallel than sequentially. Fan-out → process → fan-in structure.

Agent loop — keep iterating until the goal is achieved. "Modify code until tests pass" is the classic example. Calling the model repeatedly inside a while loop. See Agent Loop in the Agentic AI category.

Pipeline — extended chaining. Multiple stages execute in fixed order, each stage using different tools and models. Similar concept to CI/CD.

Pattern Selection Criteria

Simple Q&A → single call is enough.

Quality-critical generation → chaining (generate → validate → fix).

Diverse input types → routing.

Independent bulk processing → parallelization.

Uncertain results, iterative improvement → agent loop.

Complex multi-stage processing → pipeline.

In practice, these patterns are combined. Claude Code uses chaining and parallelization within an agent loop.