AGENT ENGINEERING

Orchestrating AI agents in financial systems

An agent may interpret a request and choose a tool. The wider financial process still needs to know which case is running, what has been approved, whether an external action completed and what should happen next. Orchestration makes those responsibilities explicit.

Key ideas

  • Choose the process and execution requirements before choosing a framework.
  • Agent reasoning, durable workflow state and business records can belong to different layers.
  • Checkpoints and approvals do not by themselves prevent repeated external effects.

Define the process before the agent topology

Describe the event that starts the work, the possible states, the tools available and the condition that ends the run. Include the paths for missing evidence, a rejected proposal and a service that fails to return a clear outcome.

Anthropic distinguishes predefined workflows from agents that dynamically direct their own work. That distinction helps decide how much choice a model needs. A known sequence may fit a conventional application workflow with a few model calls; open-ended investigation may benefit from a more flexible agent loop.

Read: Anthropic: Building effective agents

Multiple agents are useful when separate tasks, contexts or tools make the division worthwhile. They also create more intermediate results to coordinate and evaluate. Start with the roles the task requires and make the reason for each additional agent explicit.

Separate three kinds of responsibility

  1. 01

    Business workflow

    Own the case state, waiting periods, review decisions and the conditions for moving to the next step.

  2. 02

    Agent and tools

    Own the permitted reasoning task, retrieved context, tool proposals and the conditions for stopping or handing off.

  3. 03

    Authoritative services

    Enforce account permissions and business rules, perform accepted operations and provide evidence of their outcome.

These responsibilities can live in one application or in connected systems. The important design decision is where each responsibility is authoritative. A conversational transcript is useful context, but the case state and accepted action record need a defined owner.

Choose tools for the layer they need to own

The following is our selection guide, not a benchmark or a universal ranking. Evaluate the selected versions, deployment model and operating requirements for the actual project.

Starting points for an orchestration design
ApproachWhen to evaluate itStill define explicitly
Application codeA short, understood sequence with a small number of model callsState persistence, permissions, observability and error handling
LangGraphAgent paths need explicit graph state, checkpoints and pause/resume behaviorPersistent storage, external effects and business authorization
n8nThe task centers on connected application steps and reviewed AI tool useCredentials, review ownership, execution behavior and the system of record
TemporalA business workflow must preserve execution progress across waits and failuresActivity boundaries, retry behavior and idempotency of external writes

LangGraph documents state persistence and interrupts for external input. n8n documents review gates for selected AI tools. Temporal documents workflow execution history and Activities for external work. These features address different parts of the design.

Read: LangGraph: Persistence · LangGraph: Interrupts and side effects · n8n: Human review for AI tools · Temporal: Workflows · Temporal: Activities and retry behavior

A layered implementation may use an agent graph inside a longer business workflow. That adds another operating boundary, so use it when the workload justifies the separation. An existing workflow platform may already provide part of the required coordination.

Approve an exact proposed action

A reviewer needs to see the operation, target record, relevant values and supporting evidence. Keep that proposal with the decision, and revalidate it if the underlying request or material parameters change.

The server performing the action should check current authorization and the approved scope. A UI confirmation or a model's statement that an action is allowed is not an authorization service. An orchestration framework does not establish these business permissions on its own.

LangGraph's interrupt behavior is worth understanding during implementation: the interrupted node runs again from its beginning when resumed. External effects placed around that boundary need deliberate handling. Review the documented restart and side-effect behavior before connecting a write operation.

Read: LangGraph: Interrupts and side effects

Recover from an uncertain external outcome

Use a stable operation identifier and a documented deduplication or idempotency mechanism where the service supports one. Retain the accepted result and check authoritative status when the outcome is uncertain. If the service cannot establish the outcome safely, route the case for investigation.

Temporal recommends idempotent Activities so retries can avoid duplicate side effects. LangGraph also warns that retries may repeat a committed write. Durable state and approval gates remain useful, but neither automatically makes an external operation occur only once.

Read: Temporal: Activities and retry behavior · LangGraph: Interrupts and side effects

Separate an approval timeout from an execution timeout. In the first case, work may still be waiting for a decision. In the second, a connected system may already have acted. Those situations need different next steps and different information for the operator.

Evaluate the whole path

  • Does the workflow resume with the correct case and decision after a restart?
  • Can a user with a different role access or approve the same action?
  • What happens when a reviewer rejects the proposal or its parameters change?
  • Can a repeated event or a lost response produce an unintended duplicate?
  • Do step limits, cost limits and unresolved outcomes lead to an understandable handoff?

For a first delivery, choose one workflow and test its normal, waiting, rejection and recovery paths. Record the evidence and operating requirements alongside the framework decision. At Sfxdx, we use that process to select the orchestration approach around the financial task and the client's existing environment.

Sources and further reading

Primary references checked on 15 November 2026. The implementation examples and selection guidance are Sfxdx's engineering analysis.

Explore the workflow

The linked capability example uses synthetic data and local rules to show the interaction pattern.

Explore the orchestration example ↗
Workflow orchestration ↗Private & sovereign AI ↗AI inside your product ↗