← Back to all posts

Harnessing Agentic AI Systems: Orchestrator-Worker Pattern

Problem 12 of 15: dividing a workflow across agents. The Orchestrator-Worker pattern, the God Agent anti-pattern, the Sprint Contracts frontier — one table, a short discussion, the key insight, and the important references.

harnesspattern-languageagentic-aiseriesorchestrationdelegationmulti-agent

Problem 12 of 15 in the Harnessing Agentic AI Systems series — read the index for the framing. Previous: Dynamic Tool Discovery Pattern · Next: Blackboard Pattern.

The Problem — Dividing a workflow across agents

No single context can hold a whole workflow. The work must be divided, delegated, and handed off — and the frontier has added the contract that makes delegation safe (F2).

Field P13 — Orchestrator-Worker (pattern) A11 — The God Agent (anti-pattern) F2 — Sprint Contracts (frontier)
Forces / Smell Context limits vs a single plan; specialists vs oversight; termination vs conversation. One massive agent, immense prompt, every phase in one context window. Specification vs latitude; criteria vs freedom.
Solution / Anti-solution Direct traffic using a highly capable central agent that delegates atomic sub-tasks to smaller, faster, specialized agents. "One agent, all context, total control." Before each sprint, generator and evaluator negotiate what "done" looks like — and how success will be verified — before any code is written.
Consequences / Failure Bounded contexts, parallelizable work, a plan-then-execute shape. Context avalanche and lost-in-the-middle degradation by construction — the window degrades as it fills. Ambiguity resolved at the moment of maximum leverage; the agent defines the verifier before the artifact.
Tradeoffs / Refactoring The orchestrator is a single point of failure; spec errors cascade; handoffs can lose state; termination must be a system property. P13 with decomposition and handoff: one feature at a time, structured artifacts between sessions. Negotiation overhead — bounded by keeping the contract to a sprint-sized chunk.
Evidence AutoGen (paper); Anthropic's planner-generator-evaluator (post); DeepSeek's subagent registry (DeepSeek teardown). CrewAI's crews argument (docs). Anthropic's harness (post).
Related Refactoring for A11; composes with P16 and F2. Is the absence of P13; composes A4 and A8. Composes with P13 and F1.

Discussion

Delegation is how a system outgrows one context window: the orchestrator curates a delegation tree instead of drowning in context. The single point of failure is the orchestrator — the spec must stay high-level because errors cascade downstream, and termination must be a system property, not the conversation's. The god agent fails by construction; F2 makes each delegation safe by negotiating "done" before the work exists.

Key Insight

No single context can hold a workflow. Delegation is how systems scale, the spec must stay high-level because errors cascade, and termination is a system property, not a conversation's. One window holding every phase degrades exactly as the middle fills.

References

AutoGen (arXiv:2308.08155); CrewAI crews (docs); Anthropic, Harness design for long-running application development (post); archive: DeepSeek teardown, harness canon.