Agent Runtime

The shared agent model is queue and event-log based.

Core rule

Agents do not mutate provider message lists directly. Inbound input becomes a command, execution emits events, tool calls and tool results are preserved, and projections derive UI or channel output.

Why

This model keeps replay, compaction, multiple agents, delayed execution, and channel delivery from fighting over a hidden mutable transcript.

Runtime shape

inbound message
  -> command queue
  -> claimed turn
  -> model/tool loop
  -> durable events
  -> UI/channel projections

Product implication

Products may choose different storage backends and transports, but they should not bypass the runtime model by appending directly to a model-facing transcript.