AI agent orchestration: a production guide.

One orchestrator agent decomposes the task and dispatches subtasks to worker agents. Each worker has a narrow tool set and a narrow context window. Results stream back to the orchestrator for synthesis.

Ships well for: document processing, multi-source research, code refactors across many files.

Fails when: subtasks depend on each other in non-obvious ways. The orchestrator can't see the cross-cut.

A planner agent writes a step-by-step plan in structured output. An executor agent reads the plan and runs each step, with a critic agent reviewing results between steps. The plan is mutable: the critic can send the executor back for a revision.

Ships well for: long-running engineering tasks, multi-step API automations, anything where reasoning quality matters more than latency.

Fails when: the planner over-commits to an early plan and the critic doesn't catch the drift. Mitigation: make plans cheap to throw away.

N agents run the same task in parallel with different prompts, models, or temperatures. A judge agent picks the winner. Expensive but high-quality on tasks with a clear quality signal.

Ships well for: code generation with test-based scoring, content generation with rubric-based judging, search with hybrid ranking.

Fails when: the judge model is the bottleneck on quality. Bad judge = expensive averaging machine.

The agent runs autonomously until it hits a confidence floor or a side-effect threshold, then queues for human review. Human approves, edits, or rejects. The handoff is logged and feeds back into the eval suite.

Ships well for: regulated workflows (finance, healthcare), customer-facing communications, anything with reversal cost.

Fails when: the queue becomes the bottleneck and humans rubber-stamp without reading. Mitigation: random sampling for quality control.

• Fully autonomous open-ended loops. The "give the agent a goal and let it run" pattern from 2024. Cost-explodes, hallucinates plans, no audit trail. Replaced by planner-executor with a step cap.
• Agent-as-tool inside another agent. Looks elegant, debugging is brutal. Replaced by explicit orchestrator-worker boundaries.

• Step cap. Every agent loop has a hard step ceiling. No exceptions.
• Cost cap. Per-run token ceiling. Cheap-model fallback on context overflow.
• Audit log. Every step, every tool call, every model response logged with prompt, output, cost, latency.
• Eval suite. Regression suite that runs in CI. Catches drift before users do.
• Kill switch. A flag your on-call can flip to stop all agent traffic without a deploy.
• Reviewer queue. Side-effect tools route through a human queue at minimum until the eval suite has caught two months of drift.

Pick the smallest pattern that solves your job. Orchestrator-worker for most document and research jobs. Planner-executor for engineering. Swarm only when you have a clean quality signal. Supervised handoff for anything reversible-expensive. If you have a stuck multi-agent system, AI Production Systems will tell you whether the pattern is wrong or the implementation is.