What ICLR 2026 Taught Us About Multi-Agent Failures

Author here. I went through ICLR 2026 accepted papers looking for work relevant to multi-agent production problems. Found 14 papers clustered around 5 issues: latency (sequential API calls), token costs, error cascades, brittle topologies, and observability.

A few highlights: - Speculative Actions: parallel API execution, ~30% speedup - KVComm: share KV pairs instead of text, 30% of layers gets near-full performance - DoVer: intervention-driven debugging that flips 28% of failures to successes

Happy to discuss any of the papers or the framing. The decision matrix at the end maps each problem to a starting paper.