An archive of decisions, failures, changes, and shipped work. The record stays useful by keeping the sequence—and the limits—visible.
What changes when a standardized agent interface can operate physical equipment: capability, authority, observation, recovery, and evidence become part of the execution contract.
A reconstructed failure reflection on why AI-assisted work needs exact source identity, bounded verification, and evidence that matches each transition.
Why dynamic model routing turns model selection into execution policy, and what evidence, state, fallback, and evaluation controls it needs before governing real work.
Why agent architecture becomes easier to replace, debug, and govern when reusable execution and host-application authority have separate owners.
A model-churn test for agent architecture: micro-orchestration may decay as models improve, while context structure, tools, boundaries, state, and scaffolding retain value.
Why agent regression testing has to compare the exact changed candidate against a trusted baseline on the behaviors the workload cannot afford to lose.
Why a stable agent name is not enough to identify the exact configuration that evaluation, debugging, reproduction, and rollback actually apply to.
Why capable agents need enforceable limits on execution, network reach, credentials, identities, tools, approvals, persistence, and revocation.
A May 2026 Liminal Tarot paywall failure showed why a build instrumented to expose a runtime defect was not yet the artifact that should cross the release boundary.