Designing an audit agent for reviewable work

The problem

An on-call workflow needs more than an stateless agentic system that triaged and investigated the issues before reporting the answer in slack. The people reviewing it need to understand what happened, what information informed a decision, and where the work goes next. An AI-assisted workflow has to make that context available instead of leaving it inside a transient conversation.

To solve for this during an engagement, I designed an audit-agent proof of concept to explore how an agent-enabled workflow could make decisions, evidence, and handoffs easier to trace.

The design

I separated the coordination of workflow steps from the audit context those steps need. The orchestration could direct the work while the context preserved information for later review. That separation made the responsibility of each part easier to describe and gave the handoff a defined place in the design. Agents logged idempotent events from start-to-finish in a snowflake table; and a great side-effect of this is it also exposed operational gaps in the initial process helping self-heal while we were developing the proof of concept.

The practical question was not just whether an agent could complete a task. It was whether someone could follow the work and understand the evidence and decisions associated with it without getting stuck.

What this work shows

This was a proof of concept, not a production deployment. It is an example of a hybrid agentic system and how I approach developing an AI workflow: start with the work a person needs to review, make the relevant context explicit, and identify what still needs to be resolved before an operating team takes responsibility.