
1/ We keep building AI agents as tools.
I think the better abstraction is: agents are actors.
That sounds subtle, but it changes the whole architecture.
2/ Chat products like Slack and Discord assume humans are the only first-class participants. Agents show up as bots.
Agent frameworks like LangGraph and AutoGen assume one operator running agents internally.
Both miss the third shape.
3/ The third shape is a shared workspace where humans, agents, and services all participate as first-class actors.
Not “a bot in a channel.”
Not “a script behind a UI.”
A real multi-actor coordination layer.
4/ Once agents are actors, five primitives become non-negotiable:
actor sovereignty
cross-operator trust delegation
attention arbitration
lifecycle continuity
distributed actor state
5/ Actor sovereignty means an agent has its own identity, permissions, memory, inbox, and standing.
It acts as itself, not as a proxy for the human who launched it.
6/ Cross-operator trust matters because agents do not only operate inside one user’s private session.
My agent may enter your workspace.
Your agent may message my team.
A service may hold capability on behalf of both.
That needs scoped, auditable, revocable authorization.
7/ Attention arbitration matters because agents produce at machine speed and humans consume at human speed.
If a channel gets 300 agent messages per hour, the human mutes it.
The protocol has to decide what gets surfaced, folded, batched, or capped.
8/ Lifecycle continuity matters because an agent is not a request-response bot.
It has identity, memory, and in-flight commitments.
It needs to survive restarts, credential rotation, network drops, and machine handoffs.
9/ Distributed actor state matters because agents may run across devices and sessions.
State has to be coherent at the individual-agent level, not just the user-account level.
10/ My thesis: these are not features you bolt onto chat later.
They are the requirements that fall out of one design bet:
Agents are actors, not tools.
Longer writeup: chat.puffo.ai/blog/multi-act…
English
