Protecting Systems of Record in Agent Workflows

Assign every system a process role

For each application or data source, state whether it originates a request, supplies context, receives an approved action, or owns the final outcome. A system can play more than one role, but the role should be explicit at each step.

Separate read access, proposed action, and write authority

Reading a supplier record is different from changing bank details. Preparing an entry is different from posting it. Define these permissions separately and keep high-impact actions behind stronger validation or human authority.

Design reconciliation and failure recovery

The workflow should know whether an action succeeded, failed, timed out, or produced an ambiguous response. Record an idempotency reference where practical, retain the evidence needed to investigate, and avoid repeating an uncertain action automatically.

Specify a write contract for every system action

For each create or update action, define the target object, required fields, authority, validation, expected prior state, idempotency behavior, and success response. State whether the workflow may execute, prepare for approval, or only recommend. A connector that can technically perform an action does not establish that the process is authorized to use it.

Use identifiers from the owning system rather than matching on names or free text when possible. Validate that the target record still has the expected state immediately before a material change. If the state has changed, route the case for review instead of forcing an update based on stale context.

Make reconciliation part of the workflow

A successful API response is not always the final business outcome. Confirm that the owning system contains the intended record, status, and references. Where processing is asynchronous, define how the workflow checks completion and how long it waits before escalating. Preserve correlation identifiers across systems so support teams can trace one case without relying on timestamps alone.

Assign responsibility for ambiguous outcomes, duplicates, and differences between workflow state and system state. Reconciliation should identify which record is authoritative, who may correct it, and whether downstream steps must pause. Automating the initial write without designing reconciliation transfers work to operations rather than completing the process.