Home/ Incident database/multi-agent-recursive-loop-47k-2026-03
Financial loss Runaway loop verified

Four-agent LangChain loop runs 11 days and burns $47,000 while dashboards look healthy

Teja Kusireddy team (multi-agent research system) · Mar 23, 2026 · General agent

What happened

A multi-agent research system managed by Teja Kusireddy's team generated a $47,000 invoice due to a recursive communication loop. Two agents exchanged thousands of messages over eleven days without producing useful output, while dashboards showed healthy activity and normal latency -- the billing damage was surfaced only after the invoice arrived. The post-mortem identified two root causes: no per-agent budget ceiling, and no enforcement mechanism that would have terminated the session before the next API call.

$47K
Loss / impact
Mar 23
2026
Runaway
Failure mode
General
Agent type

Causal vector

Unbounded multi-agent communication loop with no token or cost ceiling and no pre-execution enforcement -- only asynchronous monitoring dashboards

Source

Reported by Waxell (citing primary post-mortem by Teja Kusireddy). Verified against the primary report.

Read the original report ↗

How a spend firewall would have helped

This is the textbook runaway loop in a multi-agent setting. A velocity cap on calls per minute per agent and a daily total cap return BLOCKED on the call that would push spend over the ceiling -- stopping the loop cold at the infrastructure layer, not as a dashboard alert someone may or may not act on.

The six rule types that contain this class of failure

Per-transaction cap

Any single spend above your ceiling is BLOCKED before it moves.

Daily total

Cumulative spend across all agent calls, bounded per day.

Velocity limit

Stops runaway retry loops — the #1 cause of overnight losses.

Merchant allowlist

Only approved destinations can ever receive funds.

Category rules

Flag high-risk classes (crypto, infra, refunds) for review.

Approval threshold

Above a value, the action waits for a human.

Related incidents

Don't be the next entry

Every incident in this database is the result of trusting a prompt, a provider cap, or a human review cycle. sipi.bot replaces all three with one deterministic call. 85 documented failures, one control.