Bound an agent workflow and its handoffs
Choose where a workflow needs model judgment, control its external actions, and preserve a clear owner for failures and human escalation.
What you'll be able to do
- Choose between a fixed workflow and dynamic tool selection for a customer task
- Define tool permissions, termination conditions and escalation outcomes
- Explain the coordination cost of adding another agent
Before this: enterprise-ai-architecture
Decide which part of the customer's task requires judgment. A warranty workflow may have predictable steps for fetching an order and checking the policy, followed by an open-ended explanation. Represent the predictable steps explicitly. Give the model the narrower decisions for which flexibility is useful.
Design the execution boundary
The model can propose a tool call. The runtime validates the arguments, checks the user's authority, runs the operation and records the result. Keep credentials in the service responsible for that operation. A description of a tool's purpose helps the model choose it, while application code enforces its actual permissions.
The runtime checks each operation and retains the result needed to recover or explain an action.
For a fictional delivery-exception assistant, provide read-only tools for order status and carrier notices. Let it recommend an option to an employee. Adding a shipment-changing tool requires a customer-approved authority policy, parameter validation, confirmation behavior and replay protection.
Set termination conditions
Specify a step limit, elapsed-time budget and cost budget appropriate to the task. Also define useful stopping outcomes. The request can complete, need clarification, fail because a dependency is unavailable, or require human review. Preserve enough state for the next owner to continue without repeating an uncertain write.
Test a tool returning incomplete data, a repeated tool request, a conflicting instruction in a carrier notice and an API outage. Inspect both the final answer and the action trace.
Add another agent for a defined purpose
A second agent may isolate research or review that can proceed independently. Define the input, output and owner of its work. Include its latency, cost and failure path in the evaluation. Give a single component responsibility for a consequential write so independent proposals cannot produce duplicate actions.
Keep the simplest design that meets the task's measured needs. Compare it with a fixed workflow on the same representative cases before expanding orchestration.
Practice. Your agent repeatedly asks a carrier for a parcel that does not exist. Explain the stopping condition, user-visible outcome and trace you would need to diagnose it. Then identify the person or system that can resolve the missing identifier.
Practise this stage with ChannelPulse
Use a system-design practice answer to narrate the action boundaries, then inspect the feedback for missing failure handling. Add your project details to Profile so preparation can use your experience. Practise system design. Prepare your profile.
Reading and attribution
bicced's agent guide is licensed under CC BY 4.0. Tech Mahato's enterprise guide informs the customer integration context. The delivery-exception exercise and explanation are original ChannelPulse material.