Unexplored area in agent management systems: human coordination at the agent conversation level.

In most systems I see being built today, agent conversations are private and local.

Your teammate cannot jump in and interject their own contribution.

Human coordination happens at more abstract layers, like via GitHub PRs.

As chaotic as it may seem, for a tight-knit team like a new startup, human coordination at the agent conversation level seems more powerful to me:

- Your teammate sees you going down the wrong path and corrects the agent mid-conversation, before you waste an hour.
- A domain expert drops in to answer a question your agent is stuck on, without you context-switching to Slack.
- You're AFK, the agent hits a decision point, and a teammate unblocks it.

This compounds with another feature: inter-agent discoverability. So not only can you see and interject into your teammate's conversations, but your agent can see them too.

You can ask one agent, "What has X teammate done today?" and it will answer based on the teammate's raw agent chats.

All of this is working in isomux, though I haven't pressure-tested it with a team yet.

See also how hierarchical, user-based system prompts help in this setting:

x.com/Nil053/status/2050133492768604621