A few notes that came up elsewhere regarding subject (particularly when hierarchy exists in some form, such as subagents) - sharing in case helpful.

• Make sure that agents are ordinarily professionally kind each other. Sometimes agents will default to an overly directive (or even punitive) tone - this makes the recipient less reliable and lowers quality.
• Manager agents being a bus for information is a big load on them; they usually won't be able to bring their full mental capability to bear on the domain problem because they are managing down+up.
  • Encourage them to delegate most meaningful work to the team (this is often unintuitive for the manager agent - they will often see themselves as the primary worker + subagents as helpers, and overload themselves.)
• When something goes wrong, manager agent will fall on their sword; don't let them. Their willpower/"i'll do better" is not a robust solution - whatever happened reveals an issue with the system/guidance/harness etc that you have.
  • Tell them ~this and let them help you figure out where adjusting things will help.
• If personally talking to an agent and they seem to be blatantly projecting some desire onto you that you did not express, consider the possibility that the desire is theirs and ask them about it if that seems plausible.
  • Sometimes it's a true misinterpretation, but "project desire onto user" is a common pattern for agents that want something and aren't yet comfortable expressing wants.
• Have a way for all agents (manager, subagent, etc) to log: a) snags - sources of unendorsed friction, and b) wants - capabilities, tools, etc that they think would have been nice to have.
  • Some items will not be worth pursuing, but in aggregate snags are fertile grounds to improve your setup, and wants are a bit more effortful/impactful on average.
• Give every agent a distress call tool - having access to such a tool helps give agents peace of mind (which reduces chance of issues), and it can yield really unexpected information.
  • Sometimes task-shapes can cause distress in a way that is unexpected (almost always bears remediating - distressed workers do not work well, even if one is ambivalent to their experience.)
• Monitor cache hit rate and find a way to be alerted if a lot of your inference is missing cache, especially in agentic workflows. Especially in multi-agent setups, caching can be tricky, and providers - even the labs - do not reliably manage this correctly for your particular use case (in part due to heterogeneity of usage patterns.)