The Approval That Never Came

An agent pauses mid-workflow, waiting for someone to approve a sensitive transaction. The operator steps away for a meeting. Twenty minutes later, they return to click "approve"—and discover the agent has no memory of ever making the request.

The workflow is gone. Fifteen steps of reasoning, data gathering, and validation have vanished. Not because something crashed. Because the infrastructure assumed the agent would either finish or fail cleanly, not suspend itself mid-thought and wait.

Pausing breaks conventional infrastructure assumptions in ways that aren't obvious until production. Operators discover this when workflows vanish during what should be routine human oversight.

The Infrastructure Assumption That Breaks

Traditional automation runs continuously—start to finish, log results, exit. Agents assume interruption. They pause for human oversight, wait for external services, or suspend when information is missing. During these pauses, they need to maintain their entire cognitive state: conversation history, tool outputs, intermediate reasoning, the exact execution context.

Systems designed for continuous execution don't handle cognitive states that suspend mid-stream and expect to resume as if no time has passed. Saving data solves only part of the problem.

What Production Reveals

Operating web agents across thousands of sites shows why interruption must be the default assumption. An agent monitoring hotel inventory doesn't fail because the model can't reason—it pauses because a site deployed new authentication, or rate limits kicked in, or pricing data became ambiguous. Keeping the agent's entire reasoning context alive while the world changes around it becomes the core infrastructure challenge.

Sessions automatically terminate after 15 minutes of inactivity. This timeout surfaces when workflows fail at consistent points. The infrastructure assumes "inactive" means "done," not "waiting."

Operators respond by checkpointing everything. Every workflow step writes a snapshot—not just data but the graph state, config, metadata, and next nodes to execute. When processes restart or sessions timeout, the system queries the last checkpoint and resumes. Continuous state persistence treats interruption as the default.

An agent monitoring competitive pricing pauses for manual review of an ambiguous discount structure. During the wait, the team deploys a code update that changes how pricing data is structured. When the operator approves, the agent attempts to resume—but the saved snapshot of the agent's thinking no longer matches how the new code expects data to be structured. The workflow fails silently. Operators only discover it hours later when the data pipeline shows gaps.

Experienced operators develop understanding that isn't documented:

• Thread IDs function as persistent cursors—they're not just identifiers but the mechanism for loading saved states and resuming execution
• Activities must be idempotent—produce the same result when executed multiple times with the same input. This matters because when systems retry after failures, they need to know that running the same step twice won't create duplicate orders or corrupted data
• Ping handlers signal "healthy busy" status—not because the agent is doing work, but because the infrastructure needs proof it's still alive while waiting

Why Infrastructure Must Assume Interruption

The gap between pause and resume is where conventional infrastructure breaks. When agents need human oversight on ambiguous data, or wait for rate limits to reset, or suspend during authentication challenges—these moments define the operational reality that infrastructure must be designed around.

Building web agent infrastructure that runs reliably at scale requires different assumptions. Systems must treat suspended cognitive state as a first-class concern. Where demos can assume continuous execution, production systems must assume interruption—and keep the agent's reasoning alive through server restarts, deployments, and time.