When a supplier changes payment terms during a purchase, finance may need to revisit an approval it has already given. Legal may need to review a revised agreement. Procurement needs to know which work can continue while those decisions are pending.
A workflow can route each request to the right team. As several agents begin working on the purchase, it also needs to track which decisions still apply, resolve dependencies, and confirm that updates reached the right systems.
AI agent orchestration gives multiple agents a way to coordinate that work under defined rules. When the agreement changes, the process must identify which approvals still hold and what needs another review.
TL;DR
- The IDC InfoBrief sponsored by Leah reports that 50.6% of organizations relied on basic orchestration, while 7% reported advanced orchestration with multi-agent collaboration and monitoring.
- Respondents planned roughly sixfold growth in agent deployments between January 2026 and January 2027. That is a plan, not a measured outcome.
- Coordinating enterprise agents requires a current record of the work, clear dependencies, action permissions, approval checks, and recovery when systems fail.
- Start with one process and test changed inputs, expired approvals, and interrupted actions before expanding its scope.
The orchestration gap in the IDC research
The September 2026 IDC InfoBrief, The Three Pillars of the Agentic Enterprise, examines how organizations are deploying and coordinating AI agents. Its orchestration findings on page 6 show a gap between expansion plans and operating maturity:
- 50.6% relied on basic orchestration, described as rule-based or workflow-based handoffs. The accompanying chart rounds this to 51%.
- 7% reported advanced orchestration, described as multi-agent collaboration with monitoring.
- 58.3% ranked agent orchestration tools and capabilities among the improvements most needed to address deployment challenges.
The report also shows deployments rising from 2.7 agents in January 2026 to a planned 15.6 in January 2027, presented as roughly sixfold growth. These figures describe the study's reported baseline and intentions. They do not establish that the planned expansion has happened or that adding agents improves business performance.
The practical question is how the organization will coordinate a larger set of agents when their work depends on shared information and decisions.
What is AI agent orchestration?
AI agent orchestration coordinates the tasks, communication, and execution of multiple agents toward a business objective. In an enterprise process, that includes deciding which work can run, what information each agent receives, and what happens when a result changes the next step.
AI agent interoperability makes it possible for agents and systems to exchange usable information. Orchestration determines how that information advances the work.
There is a place for both fixed rules and model-directed decisions. In Building effective agents, Anthropic distinguishes predefined workflows from agents that choose their next steps dynamically, and recommends starting with the simplest approach that meets the task's needs.
A stable approval route may be best expressed in rules. A specialist agent may help interpret an unfamiliar document. The organization still defines the conditions under which either can proceed.
What changes when several agents share a process?
Consider an illustrative supplier purchase already under review. Procurement is checking the supplier record, legal is assessing the agreement, and finance is reviewing the commercial terms. Some of that work can run in parallel.
Then a revised agreement arrives with a different payment schedule.
A basic implementation might continue using the earlier finance approval because its status is still marked complete. Another might restart every review, including checks the change did not affect. Both approaches create avoidable work or exposure.
A more capable orchestration design would identify the changed terms, determine which reviews depend on them, and hold the affected downstream actions. It would send the revision and relevant context to the authorized reviewer, then record the resulting decision against the current agreement.
This example is a proposed design, not a customer result. Its value is that it gives a team something concrete to demonstrate and test. The related contracting use-case article follows the broader lifecycle; the orchestration question here is how that process stays consistent when work changes midstream.
What enterprise orchestration needs beyond a handoff
A current record of the work
Agents need to agree on the supplier, agreement version, requested outcome, and status of outstanding decisions. A shared conversation history can help, but the process also needs explicit records that distinguish a draft, a recommendation, an approval, and a completed system update.
Each record should identify its source and when it was last checked. If two agents return conflicting terms, the process needs a way to resolve the conflict before either result is used.
Dependencies that survive a change
Parallel execution is useful when tasks are independent. Finance can assess a payment schedule while legal reviews an unrelated clause. Final approval may depend on both results.
The orchestration layer needs to know which results support each subsequent action. When a source document changes, it should identify the affected dependencies and request the necessary rechecks. The unaffected work can remain valid where the organization's rules allow it.
Permission checks at the point of action
An agent that recommends revised terms may have no authority to approve them. An agent that can read a supplier record may have no authority to change it.
Define those permissions for each action. Bind approvals to the relevant document or data version, and recheck the conditions before making a consequential update. A required human decision should pause the affected action and give the reviewer enough context to decide.
Recovery that accounts for what already happened
Suppose a system accepts a supplier update but the confirmation message times out. Repeating the update immediately could create a duplicate record or repeat an action that already succeeded.
The recovery process should establish the actual state before retrying. Where that state cannot be confirmed, it should preserve the uncertainty and refer the case to an owner. Completed steps, pending steps, and actions with unknown outcomes need different treatment.
Anthropic's account of building its multi-agent research system describes the importance of persistent state, checkpoints, recovery, and production tracing. Its research workload differs from commercial operations, but the engineering lesson is useful: long-running agents need a way to resume and explain their work after interruptions.
Evidence that the whole process completed
An agent's successful response is one event within a business process. Completion might also require a validated system record, an approved agreement version, and a named owner for the next obligation.
Define those conditions before the pilot starts. Then record enough evidence to check them. The operations team should be able to identify a stalled process and understand what it needs without reconstructing several agents' conversations.
A practical path from basic workflows to coordinated execution
Begin with a recurring process whose inputs, systems, and decision owners are known. Keep the existing rules that work, and establish a baseline for completion quality, elapsed time, staff effort, and operating cost.
Next, make its state and dependencies explicit. Identify where a document revision invalidates a decision, where independent checks can run together, and which actions require approval. Add specialist agents where their contribution can be evaluated against the baseline.
Before expanding, run the process through exceptions such as these:
| Test case | Behavior to verify |
|---|---|
| Commercial terms change after approval | The affected approval is checked again before dependent work proceeds. |
| Two agents return conflicting information | The discrepancy remains visible until an authorized source or reviewer resolves it. |
| A system update returns no confirmation | The process checks whether the update succeeded before attempting it again. |
| An approver is unavailable | The case follows the organization's delegation or escalation rule and retains its decision history. |
This is an evaluation sequence, not a requirement to make every process more autonomous. A predictable task may remain a rules-based workflow throughout. Add coordination capabilities where the exceptions and dependencies justify them.
Measure the completed process
Track the share of cases that meet all completion criteria, including correct records and valid approvals. Separate agent work time, system wait time, and human review time so a faster response does not conceal a longer queue elsewhere.
Include staff correction and supervision in the cost per completed case. Record duplicate actions, missed dependencies, and failures to stop for approval. These measures help a team decide whether to expand the workflow, revise its controls, or simplify the design.
For consequential writes, compare the final system state with the approved outcome. Logs showing that a tool was called are useful for diagnosis; the destination record provides evidence of what changed.
Where Leah Maestro fits
Leah Maestro is Leah's approach to coordinating commercial work across legal, contracting, procurement, and finance. Leah describes Maestro as an agentic harness that coordinates function-level orchestrators, applies the organization's playbooks, and records execution. Its current product page describes approval before irreversible actions and visibility into active workloads.
The skill-based orchestration overview explains Leah's approach to applying domain capabilities within that work.
An evaluation should use the organization's own documents, approval rules, and connected systems. Ask to see a changed-input case, an interrupted action, and the evidence that the process completed correctly. Those demonstrations make the orchestration requirements observable.
Read the full IDC study to examine the orchestration findings alongside governance and cross-functional collaboration. Then take one existing workflow and mark every point where changed information would require a new decision. Those are the places to test before adding more agents.
Frequently asked questions
How is AI agent orchestration different from workflow automation?
Workflow automation follows a defined process, which can include branching, approvals, and recovery. AI agent orchestration coordinates agents within or across that process, including their tasks, dependencies, and results. The two can work together; the useful distinction is what the implementation can coordinate and verify.
Does orchestration require fully autonomous agents?
No. An orchestrated process can combine fixed rules, agent recommendations, permitted actions, and required human decisions. The organization determines the authority for each step and the conditions under which work must pause.
When is multi-agent orchestration useful?
It is a candidate when a process benefits from distinct specialist capabilities or independent work that can run in parallel. Assess whether those benefits justify the added coordination, latency, and cost. A simpler workflow or single agent may meet the need.
What did IDC find about orchestration maturity?
The September 2026 InfoBrief reports that 50.6% of organizations relied on basic orchestration, rounded to 51% in its chart. Only 7% reported advanced orchestration with multi-agent collaboration and monitoring. The underlying study covered 410 enterprise decision-makers.
What should a team test before scaling agent workflows?
Test changed inputs, conflicting outputs, unavailable approvers, and actions whose outcomes are uncertain. Check whether the process preserves valid work, requests the right decisions, avoids duplicate actions, and leaves evidence of the final result.
