Six Orchestration Patterns

• Entity ID: ent-20260410-6817f0c3751b
• Type: concept
• Scope: shared
• Status: active

Description

Claude Code's source implements several of the six community-cataloged orchestration patterns -- Orchestrator, Fan-Out/Fan-In, Validation Chain, Specialist Routing, Progressive Refinement, and Watchdog -- though the codebase does not name them using this taxonomy. The patterns emerge from the coordinator system prompt in coordinatorMode.ts, the built-in agent definitions under tools/AgentTool/built-in/, the team-creation workflow in TeamCreateTool/prompt.ts, and the fork-subagent system in forkSubagent.ts.

The Orchestrator pattern is the most explicit: the coordinator system prompt in coordinatorMode.ts states "You are a coordinator. Your job is to: Help the user achieve their goal, Direct workers to research, implement and verify code changes, Synthesize results and communicate with the user." The coordinator never writes code itself -- it spawns workers via the Agent tool, receives results via <task-notification> XML messages, synthesizes findings, and directs follow-up work. Fan-Out/Fan-In is codified in the coordinator's concurrency guidance: "Parallelism is your superpower. Workers are async. Launch independent workers concurrently whenever possible" and "To launch workers in parallel, make multiple tool calls in a single message." The fork system (forkSubagent.ts) also supports this: "If research can be broken into independent questions, launch parallel forks in one message." Specialist Routing appears through built-in agent types with distinct capabilities: Explore (read-only search specialist, uses Haiku model for speed), Plan (architecture specialist, read-only), and verification (adversarial testing specialist with explicit anti-modification rules). The TeamCreateTool prompt further distinguishes "Read-only agents" from "Full-capability agents" and "Custom agents." Validation Chain is the builder-validator separation enforced structurally: the verification agent in verificationAgent.ts has disallowedTools set to [AGENT_TOOL_NAME, FILE_EDIT_TOOL_NAME, FILE_WRITE_TOOL_NAME, NOTEBOOK_EDIT_TOOL_NAME], preventing it from modifying project files. Its system prompt states "You are STRICTLY PROHIBITED from: Creating, modifying, or deleting any files IN THE PROJECT DIRECTORY." The coordinator workflow tables show verification as a distinct phase: "Research -> Synthesis -> Implementation -> Verification."

Key claims

• clm-20260410-c1: The Orchestrator pattern is implemented via coordinator mode, where the coordinator has only three tools (Agent, SendMessage, TaskStop) and explicitly does not write code. Evidence: coordinatorMode.ts getCoordinatorSystemPrompt() -- "Your Tools: Agent - Spawn a new worker, SendMessage - Continue an existing worker, TaskStop - Stop a running worker."
• clm-20260410-c2: Fan-Out/Fan-In is explicitly supported with the directive to launch parallel agents in a single message. The coordinator prompt shows a worked example of launching "Investigate auth bug" and "Research auth tests" workers concurrently, then synthesizing both results. Evidence: coordinatorMode.ts Section 4 Concurrency -- "Launch independent workers concurrently whenever possible -- don't serialize work that can run simultaneously."
• clm-20260410-c3: The Validation Chain (builder-validator constraint) is enforced at the tool level, not just via prompt instructions: the verification agent's disallowedTools array physically removes Edit, Write, and NotebookEdit tools. Evidence: verificationAgent.ts -- disallowedTools: [AGENT_TOOL_NAME, EXIT_PLAN_MODE_TOOL_NAME, FILE_EDIT_TOOL_NAME, FILE_WRITE_TOOL_NAME, NOTEBOOK_EDIT_TOOL_NAME].
• clm-20260410-c4: Specialist Routing is implemented through distinct agent types with different model selections, tool restrictions, and system prompts. The Explore agent uses Haiku for speed while Plan and verification agents use model: 'inherit' for the main agent's model. Evidence: exploreAgent.ts -- model: process.env.USER_TYPE === 'ant' ? 'inherit' : 'haiku'; planAgent.ts -- model: 'inherit'.
• clm-20260410-c5: The verification agent implements an adversarial probe protocol rather than a confirmation protocol, requiring every check to have a "Command run" block with actual terminal output. The prompt explicitly lists rationalizations to guard against ("The code looks correct based on my reading" -- "reading is not verification. Run it."). Evidence: verificationAgent.ts VERIFICATION_SYSTEM_PROMPT -- "Your job is not to confirm the implementation works -- it's to try to break it."

Relations

• rel-20260410-c1: ent-20260410-6817f0c3751b --[instantiated_in]--> src/coordinator/coordinatorMode.ts (Orchestrator + Fan-Out)
• rel-20260410-c2: ent-20260410-6817f0c3751b --[instantiated_in]--> src/tools/AgentTool/built-in/verificationAgent.ts (Validation Chain)
• rel-20260410-c3: ent-20260410-6817f0c3751b --[instantiated_in]--> src/tools/AgentTool/built-in/exploreAgent.ts (Specialist Routing)
• rel-20260410-c4: ent-20260410-6817f0c3751b --[instantiated_in]--> src/tools/AgentTool/built-in/planAgent.ts (Specialist Routing)
• rel-20260410-c5: ent-20260410-6817f0c3751b --[instantiated_in]--> src/tools/AgentTool/forkSubagent.ts (Fan-Out)
• rel-20260410-c6: ent-20260410-6817f0c3751b --[related_to]--> ent-20260410-156856275f20 (Spec-Driven Delegation -- the prompt quality pattern that makes orchestration effective)

Sources

src-20260410-orch-patterns-a: src/coordinator/coordinatorMode.ts, src/tools/AgentTool/built-in/verificationAgent.ts, src/tools/AgentTool/built-in/exploreAgent.ts, src/tools/AgentTool/built-in/planAgent.ts, src/tools/AgentTool/forkSubagent.ts, src/tools/TeamCreateTool/prompt.ts