Corey JamesGovernance, a runtime and collaboration: each time, I treated many agents as a distributed system held to the accountability I expect from people, with explicit state, separated duties, traceable messages and failure domains drawn on purpose.
Governance: team state in a work graph
AgentOS started from a diagnosis: long agent runs drift because the lead agent has to hold every teammate's state in its head, and past some horizon it forgets one. I moved team state into a PostgreSQL work graph.
Team state lives in a database, because an agent that must remember everyone's status eventually forgets one.
Seats from more than one vendor run in containers on more than one machine under one contract, emit signed events and carry cost per seat, and adding a vendor means adding an adapter to that contract. I assign roles to models by judgment load and cost.
Each seat gets a small, freshly rendered job. A blind spot the builders share does not pass twice, and the gatekeeper judges only what was delivered.
See Why long agent runs fail, Separation of duties for AI work and Directing AI coding agents at program scale.
A runtime: agent classes as declarations
Agent Swarm answers a different problem: every new agent tends to arrive with its own tools, permissions and messaging, and each one is a new place to fail. I made an agent class a declaration (a tool subset, a policy, a workspace and a container profile), so a new agent inherits the same safety machinery as every other with no new code. A registry tracks agents with heartbeats, a reaper and drift detection.
The envelope carries its call chain; a depth and cycle check refuses the loop.
The platform is split into planes, so the failure stays in its own.
Federation lets one swarm span several machines. See Tools are the platform and A coding agent that never holds the keys.
Collaboration: models and people in one room
Helix puts several models in one room with people. Its Control Center, on an unmodified Mattermost server, gives each model vendor its own bot seat and adds a council mode where models deliberate, challenge each other and ratify a decision.
A question fanned out, answers collected. Thrown out.
Directed routing sends a message to the participant it names.
In voice, several AI voices share one conversation without talking over each other, and all of them yield when a person speaks. The models are vendors' and open-source; the orchestration, state, messaging and duty structure are mine.