How does agent orchestration work?
An orchestration layer sits above the individual agents and handles four jobs. It routes each task to the agent best suited for it. It manages handoffs, passing context from one agent to the next so work does not restart from zero. It prevents conflicts — typically with locks or task ownership, so two agents never write the same resource at once. And it records every decision in a log a human can read. Anthropic's multi-agent research system and OpenAI's Agents SDK both formalize this as an orchestrator-worker (or manager) pattern: one coordinating agent delegates to worker agents and synthesizes their results.
Why do multiple agents need an orchestrator?
Because parallel autonomy fails without coordination. Two agents that both edit the same email template produce a mangled template. Two agents that both message the same customer produce spam. An orchestrator serializes writes, deduplicates work, and keeps every agent aligned on the same goals. This is also part of what separates agents from chatbots: a chatbot answers you, while orchestrated agents act on shared systems — which only stays safe if something coordinates them.
What does agent orchestration look like in a marketing platform?
Often, the orchestrator sits outside the platform. A team runs its own agent in an MCP client such as Claude Code or Cursor, and the platform exposes tools over the Model Context Protocol: create a segment, draft a journey, read a report. The client sequences the calls; the platform sets the limits: which actions exist and which need a confirmation step (publishing a journey, for example). The platform can also write every call to an audit log.
What should a human be able to audit?
Everything the agents changed, and why. A usable decision log records the action, the agent that took it, the inputs it relied on, and the reasoning — not just a timestamp. Human-in-the-loop oversight pairs that log with a confirmation step: the step holds actions that are hard to undo until someone confirms them, and the log is how you trace everything else afterwards. Orchestration without an auditable trail is just automation you cannot inspect.