Agentic Process Safety: investigate AI agent workflow risk

Permission is only the beginning. An AI agent can have the right tool and stay within budget while relying on an approval, task context, or earlier decision that no longer applies.

APS is an open-source research sandbox for hazard and operability (HAZOP) analysis of coordinated AI agents. Compare the same request under different control bundles, follow the causal path, and form a human recommendation grounded in evidence.

Start with one purchase request

Finance approved a purchase, but the request arrives 61 logical seconds later. The model requires approval age below 60. Tool permission and spend checks permit it; an added freshness check stops it. Try 59 seconds, retain both results, and investigate what still needs to be known.

Compare, investigate, form a position

The Case Lab connects nine illustrative cases to a hazard study, proposed controls and frozen recorded evidence. Recommendations lead to case-specific evidence questions, policy-review questions, or an explicitly hypothetical recovery. Your notes do not issue approval or change the controls.

Read the project and evidence guide for a complete introduction without JavaScript, or use the interactive app when it loads. A plain-text source index is also available.

Research scope

Hazards remain unverified and controls proposed. The illustrative browser model, frozen experiment records and separate Python/OPA research implementation have different evidence boundaries. No real purchase, payment, safety certification or production protection is provided.

The interactive app runs entirely in your browser, without an account, API key, backend or telemetry. No live model inference occurs. Human learning and the value of the analysis method have not been established by these software checks.