Agent orchestration is the coordination of several specialised AI agents within one process, deciding which agent acts at each step, what context it receives, how results pass from one agent to the next, and where a human must approve before the process continues.
Agent orchestration extends workflow orchestration, which coordinates tasks and systems, to steps performed by AI agents whose output is probabilistic. The orchestration layer therefore has to verify outputs, not only sequence them.
Four patterns cover most finance use cases, following the taxonomy Anthropic published in its December 2024 engineering guidance on building effective agents:
- Sequential chain: each agent hands its output to the next, for example intake, extraction, matching and approval
- Routing: a first agent classifies the input and sends it to the right specialist, such as invoice, credit note or supplier statement
- Parallel checks: several agents check the same document independently, for price, duplicates and bank details, and results are combined
- Orchestrator and workers: a coordinating agent breaks a task such as a month-end review into subtasks and assigns them
Errors compound across agents. If each of four chained agents is right 95% of the time, the chain is right about 81% of the time, since 0.95 to the fourth power is 0.81. Orchestration therefore needs checkpoints between steps: confidence thresholds, validation against reference data and human-in-the-loop review of exceptions. Connections to external systems increasingly rely on the Model Context Protocol.
Phacet orchestrates agents this way. The agent that can automate your accounting inbox routes each document, extraction and matching agents check it, and exceptions reach a reviewer with the reasoning attached, each step traced in the audit trail.
