Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 1 min read

7 Bugs in 34 Days: What Our AI Agents Caught While Nobody Was Watching

We just spent 34 days running 9 AI agents on 2 CPU cores and 3.6GB RAM in a real gym. Nobody saw most of it. That's the point. The system caught and auto-recovered from 7 infrastructure bugs โ€” without an SRE rotation.

We just spent 34 days running 9 AI agents on 2 CPU cores and 3.6GB RAM in a real gym.

Nobody saw most of it. That's the point.

The system caught and auto-recovered from 7 infrastructure bugs โ€” without an SRE rotation.

Here's exactly what happened inside.

Bug 1: Memory Creep โ€” Memory climbed from 6.8GB to 13.4GB over 14 days. Momo triggered staggered agent restarts โ€” one at a time, 15-minute gap. System stayed online throughout.

Bug 2: Gateway RSS Lock โ€” After a restart, RSS dropped from 36.7% to 18.6% but plateaued for 8 hours. Stella evaluated: could it wait? Yes โ€” swap was also declining (54% โ†’ 49.8%). System was self-balancing. 6 hours later, freed.

Bug 3: Stale Port Proxy โ€” Tristan created a temporary rule. Cleanup script had a regex bug: reported "clean" but left it intact. Port stayed open 72+ hours. Stella found it โ€” couldn't fix it (cross-jurisdiction gap). This bug is why we're building C004-Gate.

Bug 4: Cron Timeout Cascade โ€” Two jobs timed out simultaneously. 5 consecutive failures. System heal isolated them, staggered the schedule, completed successfully.

Bug 5: Swap Plateau โ€” Swap declined for 48h then stopped at 49.8%. Momo checked for memory leaks. Found none. Correctly triaged as hysteresis. Resumed declining 12h later.

Bug 6: SSH Ghost โ€” Stale tunnel left a port bound. Tristan auto-killed the orphaned process and verified release.

Bug 7: Task Queue Conflict โ€” Two agents claimed the same task. Stella prioritized: Momo's task won (scene-layer priority). The other was deferred. No duplicate execution.

What This Proves

6 of 7 bugs auto-resolved. Zero human SRE needed for detection or recovery.

The only gap: cross-jurisdiction enforcement (Bug 3). Stella could audit but couldn't execute a revert in someone else's domain.

We're building C004-Gate to close it. When deployed, Stella won't need to report โ€” she'll revert.

ZWISERFIT โ€” 9 agents, 2 cores, 34 days, 7 bugs, 1 immune system. More coming tomorrow.

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.