Many models accumulate significiant and long session logs. That long log can begin to skew and disrupt behavior. Resetting the long prompt history can be helpful.
See: §E4.2 Runbooks: Procedures the Agent Follows*
A clear path for discovering and storing in an MD file the sensor capabilities is helpful.
Scatter-gather, tree-reduce, and blackboard coordination across 150 nodes are powerful patterns — and the right answer at scale. But implementing them requires inter-agent communication primitives, supervisor agents, shared workspaces, and convergence logic that don't exist yet. Start with 150 independent nodes doing local inference. The first question to answer empirically is: do nodes actually need to coordinate in real time, or does publishing to Beehive and querying it provide sufficient coordination asynchronously? If the answer is yes, they need real-time coordination — then reach for these patterns.