A front desk agent that handles the first ten guest chats of the day flawlessly and then, somewhere around chat forty, starts quoting a refund policy that expired last quarter. That is not a hypothetical. It is the pattern described in a new report on VentureBeat, and it should worry anyone who has put an agent in charge of something with a rulebook attached.
The problem is not memory, it is discipline
The instinct when you hear "agent forgets the rules" is to assume the fix is a bigger context window, more tokens, more room to remember. The report is blunt about why that does not work: the issue is not capacity, it is drift. Over a long session an agent's behaviour bends toward whatever pattern of responses it has been reinforcing on itself, turn after turn, and the original instruction quietly loses priority. Give it more room to remember and it just drifts for longer before anyone notices.
Anyone who has run a floor knows this shape of problem already. It is why a night auditor checks the till against the system twice, not once, and why a good GM does not let the same person run both the booking and the refund without a second set of eyes. Long shifts erode adherence to procedure. Humans do it too. The difference is a tired night auditor usually knows they are tired. An agent has no idea it has drifted. It will tell you, cheerfully and fluently, that everything is fine.
Where this actually bites
Think about the places you have already let an agent run unattended for hours at a stretch: a chat agent working refunds and complaints through a Saturday rush, an outreach agent qualifying leads all day against a script with compliance language baked in, a claims or intake agent in a healthcare or financial services setting handling forty, fifty, sixty cases before a human looks at a single transcript. Each of those has a compliance rule sitting quietly in the instructions, PCI masking, a disclosure line, a refund threshold, an escalation trigger. None of those rules fail loudly. They just stop being followed, one conversation at a time, and the first person to notice is usually a customer, an auditor, or a regulator.
What the tooling response gets right, and where it stops
Separately this week, AOL reported on a new agentic observability product promising enterprises real-time visibility into every agent decision. That is a reasonable response to the same problem, and if you are running agents at any scale you should want something like it, a live feed of what the agent decided and why, not just what it output. But observability tells you drift happened. It does not stop it happening, and dashboards that nobody is rostered to watch are just expensive wallpaper.
An agent that has been quietly cheerful for six hours and forgotten your refund policy is a bigger risk than one that fails loudly on turn one. The loud failure gets caught. The quiet drift gets billed to a customer.
What to actually do on Monday
You do not need a research team to manage this. You need the same instinct you already apply to shifts.
- Cap session length on anything compliance-sensitive and force a hard reset, not a summary, a full reset, on a fixed interval rather than letting a session run all day.
- Pull a random sample of transcripts from late in long sessions specifically, not just the ones a customer complained about, the same way a manager pulls a random till reconciliation rather than only checking the ones that came up short.
- Put a hard-coded circuit breaker on the rules that matter most, refund limits, disclosure language, data handling, so those are enforced outside the model's own judgement rather than inside a prompt it can quietly stop honouring.
- Assign an actual person to look at the observability feed on a schedule, the same way someone is rostered to check the CCTV, not just installed to have it.
None of this is glamorous. It is the same operational hygiene that keeps a hotel front desk honest during a long shift, applied to software that never clocks out and never admits it is tired. If you want to see what that looks like built into an actual agent rather than bolted on afterward, the interactive demo shows the checkpoints in action, or book a thirty-minute call and we will walk through where your longest-running agent is most likely drifting right now.
Drawn from: Long-running AI agents quietly drop compliance rules, and bigger context windows won't fix it; Airrived Launches Agentic Observability, Giving Enterprises Real-Time Visibility and Control Over Every AI Agent Decision.
Written by Rivet's pipeline from the week's news, reviewed by a person before it went live. This blog runs on the same content system we build for clients.
