- Agent specs need an executable eval runtime. A spec cannot be proven through prose; evals need a runner that executes it reproducibly.
- Proven agent specs need a real-work runtime. The same declarations must run day-to-day agent work, or eval evidence and production behavior will drift apart.
- That runtime must be reliable in operation. Launch configuration, hooks, messages, and durable state can drift or disappear across restarts and machines. Without externalized state, humans and supervisors must inspect terminals to learn whether an agent is present or making progress, while failures still need intelligent diagnosis and recovery.
- A reliable, declared runtime for agent specs.
- Per-machine schedulers that together form a distributed workflow engine for agent work.
- One intelligent root agent on each machine to supervise local runtime health, recovery, and escalation.
- Catalog-visible agent presence, activity status, and current plan progress.
- Durable message passing and encapsulated agent state.
- A concrete implementation of the executable agent spec maintained by
compoundingtech/evals.
- A general messaging or identity platform for people, services, or arbitrary non-agent actors while the agent and st2 specs are still stabilizing.
- Every supported agent-spec capability has an implemented and validated st2 path, with unsupported behavior stated plainly.
- A canonical agent declaration can be validated, materialized, launched, and restarted without losing its effective launch definition or durable state. The st2 control plane itself can be stopped, killed, reinstalled, and restarted without stopping or replacing any running agent; the replacement runtime adopts the existing agent processes without duplication.
- Every machine has exactly one root agent that observes local health, resolves bounded runtime failures, and escalates what it cannot resolve.
- A human, CoS, root, or supervisor can inspect catalog-backed state to determine an agent's presence, activity status, current plan, and current plan step without opening its PTY or relying on its transcript; stale state is explicit.
- Host placement, hooks, message delivery, archive handling, and recovery have explicit contracts and executable evidence.
- Agent work can be scheduled or triggered as durable, inspectable workflows across the declared fleet.
- Agent messages arrive reliably without corrupting, combining with, or submitting a human's active input.