A three-agent workflow degraded for months before a reviewer spotted it: latency and error rates stayed normal, but a summarizer dropped one field and a downstream agent made a confident wrong recommendation.
The report argues agents recreate microservices’ visibility problem faster and at larger scale because teams can spin them up in an afternoon, often without inventory, ownership or end-to-end tracing.
1,600 execution traces across seven multi-agent frameworks produced a taxonomy of 14 recurring failure modes in MAST, with most problems tied to system specification and agent coordination rather than model quality.
The proposed fixes borrow from microservices: agent registries, correlation-ID tracing via OpenTelemetry GenAI spans, prompt versioning, contract tests and policy-layer tool access instead of relying on system prompts.
Cost makes the risk broader than classic outages: looping or drifting agents can quietly erode output quality and inflate unit economics, leaving product owners to discover the damage at billing time.