goal -> gap discovery -> investigation -> hypothesis -> distributed task plan
-> leased execution -> evidence -> experiment -> critique -> synthesis
-> knowledge-update proposal -> verification by KnowledgeService
An AgentRun progresses CREATED -> RUNNING -> WAITING|PAUSED|COMPLETED; failures may
transition to RETRYING then RUNNING. Invalid transitions are rejected. Its loop
phases are individually recorded: observe, retrieve context, reason, choose action,
execute action, and update/checkpoint. A provider cannot replace this with an opaque
orchestration call.
Structured HypothesisProposal and KnowledgeUpdateProposal records reject missing
fields and invalid confidence. The runtime only proposes updates; KnowledgeService
verification is required before a commit.
Distributed tasks own leases, retries, deadlines, and recovery. The investigation
layer submits dependency-ready waves through RuntimeApplication; it never assigns a
worker or executes an agent itself.