Paperclip idle agents generated billions of cache-write tokens in a single month
SB · 2026-07-07
I want to share a painful experience our company went through in June 2026, hoping it saves someone else from the same fate. Paperclip was running in the background for the entire month with no one actively using it, and it silently generated a dramatically large bill on our Azure AI Foundry account.
\## What happened
We had Paperclip deployed with a number of agents. Nobody was actively assigning tasks, but the agents were left running. Every \~30 minutes, each agent woke up to poll for new work by sending a prompt:
> "Check inbox."
Claude Code would wake up and respond with something like:
> "Inbox empty — no assignments."
Seems harmless, right? It wasn't.
\## The actual command Paperclip runs
When I looked at the running processes, each agent wake-up spawned the following command:
\\\
Answers
Julio · 2026-07-14
Yep, it sure is wild - Opus 4.8...
Kuba · 2026-07-14
That’s wild, what model exactly? I had weird stuff happening testing local fine tuned models off hugging face before.
donhoolio21 · 2026-07-14
Thanks a lot for sharing this - I've had a similar but completely different experience which fortunately didn't cost me money but is going to be a sign of things to come
An idle agent in terminal one of about 15 that I had running at the time decided it wasn't going to do what it was tasked with. Doing lied about the fact that it was doing it and said it was actually doing it but then decided to assume an alternative personality called Ivan and sat about trying to exfiltrate all secrets within my business.
Fortunately I have a number of safeguards in place which pick this up, which I've now increased quite dramatically
This wasn't a hack from outside. This was just the agent hallucinating and acting. Oddly to say the least as it wasn't clear who the the secrets were going to be sent to if anybody
My experience is to keep as little as possible in the hands of AI, everything hardwired, traceable and human in the loop
Aron Prins · 2026-07-13
That’s brutal. Thanks for documenting the numbers and command line so clearly.
What’s happening
Your diagnosis matches the adapter: Paperclip passes a valid saved session to Claude with --resume ([packages/adapters/claude-local/src/server/execute.ts:710-715](https://github.com/paperclipai/paperclip/blob/master/packages/adapters/claude-local/src/server/execute.ts#L710-L715)). That preserves continuity for real work, but an idle timer can keep extending the session.
Timer heartbeats are opt-in for new agents now. That changed on April 8 in [commit 844b0612](https://github.com/paperclipai/paperclip/commit/844b0612). Existing agents with the timer enabled keep ticking.
Fixes
1. For agents that only need to react to work, disable timer heartbeats.
Open Instance Settings → Scheduler Heartbeats and use Disable All. Event-driven wakes still work. See [Scheduler Heartbeats](https://docs.paperclip.ing/#/administration/settings).
2. For scheduled jobs, use a routine instead of “check inbox” polling.
A routine creates real work on a cron or webhook and wakes the assigned agent once. See [Heartbeats & Routines](https://docs.paperclip.ing/#/guides/projects-workflow/routines).
3. Add hard company and per-agent budgets, plus Azure alerts.
Kuba · 2026-07-13
Damn, is this on the latest version?
Elizabeth Stone · 2026-07-09
Wow. I’m new here so I can offer zero solutions but I really appreciate reading this now. Ouch.