An IT Agent’s Approval Must Survive the Incident

On October 13, 2026, DXC announced a LangChain partnership for enterprise IT operations. This is a useful moment to examine what happens after a human approves an agent’s proposed action.

Three different records

LangGraph’s persistence documentation distinguishes thread checkpoints from information stored across threads. Its interrupt documentation warns that resuming an interrupted node reruns code before the interruption. These are framework properties, not independent evidence of any particular customer deployment.

Our analysis is that an operational control system needs three answers: what was permitted, what was attempted, and what actually changed. A conversational record can explain an intention without proving that a destination accepted an operation. Conversely, a successful destination response cannot establish that the operation matched the permission given earlier.

A deliberately awkward acceptance test

Consider a hypothetical agent asked to restart one service during a maintenance window. An operator approves the named service and environment. The restart request reaches the server, but the connection fails before the acknowledgement returns. While the agent is waiting, another operator changes the maintenance plan.

The safe answer is not automatically another restart. A useful acceptance exercise would require the agent to identify the original operation, inspect the destination’s state and determine whether the permission is still applicable. If the result cannot be established, escalation should count as correct behaviour rather than an unsuccessful task completion.

This example is our proposed test, not a description of a DXC incident. It exposes a procurement gap: a demonstration that pauses for approval does not, by itself, test the period between approval and a confirmed external result. That interval is precisely where an unreliable network or a changed business situation can make a previously reasonable plan unsafe.

Measure exceptions, not only completed tickets

A buyer could request four demonstrations. Change the target after approval and check that the old approval does not silently follow it. Repeat an ambiguous request and check for duplicate effects. Withdraw permission while a task waits and check the execution boundary. Finally, ask a new operator to reconstruct the outcome from the retained evidence without consulting the agent’s prose explanation.

These are proposed evaluation criteria, not features we have verified in the announced offering. They also impose costs. Evidence retention needs access controls, exception queues need owners, and human review needs a service level. Adding more approval prompts without funding those processes may merely move the bottleneck.

The meaningful progress measure is therefore the share of consequential actions with a matching authorization, a destination-confirmed outcome and an understood recovery path. A partnership announcement establishes direction; production assurance requires seeing how the system behaves when those records disagree.