Controlling Consequential Transitions: An Execution Contract and Architecture-Review Method for Agent Workflows

Agent workflows can already combine durable execution, policy evaluation, provenance, human approval, and audit history, yet consequential actions may still depend on application conventions that are difficult to inspect and easy to bypass. This technical note treats the consequential state transition as the unit of control and asks what execution contract makes the separation of productive work, trusted evidence, authorization, and human decision points enforceable in software. It defines a seven-stage transition—Propose → Gather Evidence → Evaluate → Authorize → Execute → Observe → Commit—and derives fourteen requirements covering durable execution, trusted evidence, explicit scoped authority, enforcement at the real effect boundary, independent observation, human decisions, and reconstructable history. Existing workflow, policy, provenance, event-history, human-task, and agent-checkpointing systems provide most of the necessary primitives, but the complete evidence → policy → authority → effect → observation → commit relationship is not generally first-class across product boundaries. The note develops a reference architecture, matched-effort alternatives, failure-oriented acceptance tests, and an architecture-review method for testing whether the stronger execution contract earns its complexity.