This guide turns one proven agentic workflow into a durable production responsibility. Start interactively, then automate only the behavior, access, and failure policy you have verified.
1. Prove the workflow interactively
Use the Interactive Agent to complete the task with explicit targets and acceptance criteria. Confirm which platform context, plugins, approvals, and evidence are actually required.
Do not start by creating a broad autonomous Agent. A successful supervised run exposes missing permissions, ambiguous decisions, and unsafe external effects while a person can still steer the work.
2. Define one responsibility
Write:
- a measurable mission;
- included and excluded responsibilities;
- success criteria;
- the owner accountable for the business outcome;
- the evidence required before a run can finish.
Separate deterministic business rules into Components or Jobs. Use the Agent for interpretation, investigation, coordination, or adaptive capability selection.
3. Assign the execution identity
Select a dedicated service account with least privilege. Test the exact reads, writes, Storage folders, Secrets bindings, and plugin targets under that identity.
The Agent owner controls management. The service account controls runtime access. Neither the model nor an attached plugin broadens those permissions.
Choose the model and reasoning profile, planning policy, permission level, workspace, concurrency, run limits, and required plugins.
Start with:
- one active root run;
- conservative child-run concurrency;
- Restricted or Auto Review;
- a managed workspace;
- only the capabilities proven in the interactive workflow;
- explicit approval routing for external or destructive actions.
5. Test one direct run
Run one bounded instruction and inspect:
- admitted context and missing access;
- plan and capability activity;
- approval or clarification pauses;
- workspace outputs;
- final status and evidence;
- behavior after cancellation, retry, and continuation.
Only a terminal status plus the required durable result proves completion.
6. Add a trigger
Use Inbox first when work requires prioritization, delay, correlation, or de-duplication. Add a Schedule or Event after the task input and retry semantics are stable.
Every producer should supply a stable business correlation identifier. Repeated delivery must not repeat a non-idempotent external effect.
7. Operate by exception
Monitor blocked, action-required, failed, delayed, and capacity-limited runs. Keep an operator procedure for:
- missing permission or plugin access;
- approval delegation;
- ambiguous external outcomes;
- failed child work;
- workspace or retention problems;
- retry versus continue decisions.
Pause the Agent before materially changing its instructions, plugins, identity, workspace, or concurrency while work is active.
Last modified on September 23, 2026
