What an AI Agent Should Do When a Tool or Screen Changes

Reliable automation requires visible preconditions before any state-altering action takes place. Before an assistant clicks an action element or sends a payload to an endpoint, it should verify specific environmental markers such as expected URL paths, field labels, and element identifiers.

For example, consider a hypothetical billing workflow where an assistant prepares monthly retainer invoices in an external accounting portal. If the accounting provider moves the primary action from Approve to Archive, an agent that clicks by coordinates or ambiguous text matching might archive billable work instead of preparing drafts. When expected selectors disappear or unexpected modal dialogues appear, the correct default behavior is an immediate halt rather than an exploratory guess.

• Verify unambiguous anchor elements and input IDs before entering data.
• Halt immediately when an unexpected dialogue, schema mismatch, or relocated field occurs.
• Capture complete contextual telemetry, including screenshot state, timestamp, and target URLs.
• Generate an uncompleted task draft for team review rather than attempting alternate write actions.

Halting prevents damage, but operational processes still need completion. The system must create an informative escalation package that shows an operator what was completed, where the unexpected change occurred, and what remains pending. This allows a project manager or operations specialist to finish the task manually without starting over from scratch.

Once human staff resolve the immediate operational need, the workflow definition must be deliberately updated and tested. Automation should never adapt its own operational permissions or interface rules autonomously without human approval. Keeping configuration changes subject to team review preserves predictability and clear institutional ownership over core service operations.