Christopher Lovejoy and Saul Howard use a healthcare proof of concept to show why good model performance does not establish production readiness. Teams still need to explain every action, enforce sensitive-data boundaries, support required human approvals and evaluate behavior on representative data. Adding these requirements later can make an initially promising prototype brittle.
Saul Howard proposes a unified, immutable event log that records agent actions and the authority behind them. Event sourcing supports reconstructing prior system states and creating different views of the same history. He also makes the trade-off explicit: appending events is straightforward, while reading useful state may require projections, caching or snapshots.
Saul Howard separates the action log from schema-driven storage of sensitive data, with events holding references to protected objects. This can let developers inspect execution without receiving the underlying personal information. Access tokens checked when data is used provide a place to enforce boundaries and limit exposure, although the architecture still requires carefully implemented controls.
Christopher Lovejoy describes a common action interface for humans and model-based agents so a workflow can escalate at different points without changing downstream expectations. Shared context can be presented as a model prompt or a human-facing interface. He then connects event replay, human comparisons and controlled access to production data with evaluation, allowing teams to investigate changes while preserving data boundaries.
Christopher Lovejoy and Saul Howard recommend choosing architectural trade-offs around production constraints before rebuilding toward proof-of-concept accuracy. Their examples draw on established enterprise patterns, with the aim of making auditability, privacy, oversight and evaluation routine properties of the system.
Watch on YouTube
