Peer-edge governance

Errors don't only flow down the pipeline; they flow sideways. An agent's output forwarded into a colleague's workflow, one model's answer piped into another's context, a contaminated cache replicated into a knowledge base — peer edges couple parts of the system that governance treated as separate, so one team's (or one agent's) lax practice becomes everyone's ambient risk. The network literature is blunt about the mechanism: things spread along informal ties faster than along the org chart, and what spreads arrives carrying a peer's endorsement.

Every operator class and every model has these pathways, and they run in both signs. The reinforcing side carries failure: peer contagion between workers, agent-to-agent relays, and the monoculture case — the same model serving every case, so a single flaw repeats as correlated error across the whole caseload instead of averaging out (the mechanism behind the tens of thousands of simultaneous wrongful determinations in the documented automated-benefits failures). The inhibiting side carries checking: colleagues reviewing each other's work, and independent models cross-checking each other. Governing peer edges means dampening the first without crushing the second — a rule that suppresses informal peer review along with informal peer contagion has destroyed an asset.

Peer-edge governance therefore applies the same controls as anywhere else, on both signs: provenance survives forwarding and hand-off, agent-to-agent couplings and replication require authorization, shared stores get the strictest write rules because they have the widest read audience — and peer review is made routine and sanctioned, so the second look is a norm rather than an accusation.

The scaling warning: peer pathways multiply with headcount, integrations, and agent count. A deployment that was self-limiting as a pilot can cross the sustainability threshold purely through growth in sideways connections — which is what current agentic-governance frameworks flag when they treat inter-agent interaction as a first-class assurance surface.