TL;DR — Key Takeaways
- State management is the missing foundation: AI agents operating across lengthy workflows need to preserve progress, pause safely and resume without starting over.
- Human approvals require real infrastructure: Governance policies alone cannot manage timeouts, partial approvals, interrupted tasks or decisions that arrive hours or days later.
- Durable execution makes agents production-ready: Checkpointing, retries, recovery mechanisms and compensation workflows allow agents to survive interruptions without losing context or repeating completed actions.
You see the same bottleneck over and over with AI agents in production. This isn’t a bug, but you will most likely not see it documented in any incident ticket. The typical scenario here is: The agent takes several actions, waits for some sort of manual approval or response from another system, and nothing happens. It either hangs there in limbo, or just restarts everything from scratch once somebody responds. Teams typically call this a “governance problem,” but in actuality it’s a fairly simple issue. The agent has no place to put its current state, and no clue how to return to it after waiting.
The industry is beginning to recognize this challenge. Google, for example, recently published guidance on building long-running agents that can pause for days and resume without losing context. This reflects a broader shift in agent architecture: As agents move from short-lived interactions into real business processes, durable execution becomes a fundamental requirement. And the timing tracks: Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of this year, which means a lot of teams are about to hit this wall for the first time.
Pillar 1: Why This Feels Like New Territory
Scripting is simple. It either runs or it doesn’t. But agents are a whole different animal. They juggle tasks across systems and make decisions while running, which means they break in brand-new ways. Imagine one is in the middle of a five-step approval process. Step three is waiting on a human response, which doesn’t show up for six hours. When the answer finally lands, the agent can’t just start over or guess where it was. It has to resume where it left off. The current generation of automators was not designed to maintain an incomplete state for several hours, much less for several days. That is why the pilots run fine initially, but everything falls apart once you enter production.
State management is the foundation, but enterprise agents need more than persistence. They need durable execution: the ability to checkpoint progress, wait for events or human decisions, enforce timeouts, retry failed operations, compensate for completed actions, and resume safely after failures.
Pillar 2: Human Checks Need Actual Mechanisms
Approval via humans is the obvious choice in the realm of automation. However, “waiting for a human” can have little meaning unless there is a rule for how long to wait before moving on to other steps and what to do when the time elapses. Without those rules, the agent sits helpless. It can’t tell if the reviewer’s out sick, buried in email, or if it’s simply time to default the decision. Policy docs won’t fix this. You need a state machine that actually keeps track of the wait.
Pillar 3: Resuming Isn’t Just Hitting “Continue”
Picking up a paused workflow is rarely as simple as remembering where you left off. Has the data changed in the last few hours? Did some other system already act on old info? And if the agent submitted five items for approval, but three came back approved and two bounced, can it handle that split? Or does it treat the whole thing as one yes or no? Most systems fumble here. They start over, or they throw out the whole batch. Real-world approvals are rarely that tidy. An agent needs infrastructure that holds partial progress and knows exactly what’s been signed off and what’s still pending.
Pillar 4: Recovery Takes More Than Just Rolling Back
When a long, multi-step process blows up halfway through, people usually just roll it all back, or ignore the problem and hope nothing gets hurt downstream. That won’t be enough in production, where your agents will be handling real transactions. The thing you need is a recovery mechanism that will tell you what went wrong, what has been committed, and what to do next: retry, compensate, or pass things over to a human with all the details rather than just stack traces.
Governance Won’t Fix What’s Missing Under the Hood
Frame all of this as a governance problem, and you just end up adding more policy gates. Policies are fine, but they’re not the real fix. The problem is that the system can’t hold or resume state across a long pause. The best policy in the world means nothing if your infrastructure can’t tell the agent what happened while it was waiting. Tackle the state management issue first, and the governance conversation gets a lot saner. Now the agents can actually pause, wait, and start back up without losing track.
Plenty of companies are about to hit this wall as they scale past pilot projects. The ones ahead of the curve are already building state and recovery layers, not just piling on more policy to paper over the gap underneath.
