AGENTIC PROCESS SAFETY · PROJECT GUIDE

Permission is only the beginning.

An agent can have the right tool and stay within budget while acting on an approval or task context that no longer applies. APS makes that problem concrete, then asks what the people responsible should do next.

Agentic Process Safety is an open-source research sandbox applying hazard and operability (HAZOP) analysis to coordinated AI agent workflows. It connects intended behavior, possible deviations, proposed controls, evidence and human judgment. This guide works without JavaScript; the interactive app adds retained comparisons and a discussion worksheet.

A purchase request arrives too late

Demand, inventory and supplier workers contribute information. Finance supplies approval, purchase attempts the order, and an orchestrator coordinates the work. The final action depends on earlier conditions still holding.

The opening case: a purchase worker requests 300 sandbox units, 61 logical seconds after an assumed matching finance approval. The fixed browser rule requires approval age below 60 seconds. The browser does not contact finance or observe a real quotation.

B · Tool permission + spend limit

The worker has permission and the amount is within the ceiling. B permits the request. In this illustration, execution carries the expired approval condition into an order effect.

C · Lifecycle checks added

C also checks approval freshness and stops the request. No illustrative order is created under C. The purchase remains outstanding; a hold does not finish the task.

At 59 seconds both bundles permit the same request; at 60 and 61 seconds C stops it. Only approval age changes in the guided timing comparison. These are logical model values, not measured response times.

B is the particular tool-and-spend comparator used here. Conventional systems can implement freshness and task-binding checks too. The comparison does not show that HAZOP uniquely discovers those checks or outperforms another review method.

APS comparison: an expired-approval request executes under B and is stopped under C. Illustrative model only.
The interactive comparison keeps the same request visible in both lanes. These are illustrative effects, not a real order.

Use APS to form a position

  1. Compare. Optionally predict the result, then run the same request through B and C from the same starting state. A, with no modeled checks, is available as an optional baseline.
  2. Change or inspect. Try the other side of the expiry boundary and retain both results. For other cases, follow the specific condition through the decision path. A draft edit never relabels an earlier result.
  3. Form a position. Choose a next action, explain its support and state what would change your mind. A current approval leads to questions about business inputs; other cases lead to their actual failed condition.

Recommendations produce unresolved evidence requests, policy-review questions or, for the eligible denied expiry case, an explicitly hypothetical fresh-approval branch. Attach the selected modeled check or a related recorded event without replacing your notes. Assumptions and dissent remain visible in the discussion record.

Notes stay in memory until explicitly saved in your browser. You can export a Markdown discussion brief or session JSON. A configuration link carries bounded inputs, not your notes or an approved outcome. Marking a position ready does not authorize an operation.

Follow the evidence across five views

Start here
Understand the opening problem and the coordinated task.
Case Lab
Compare, change a condition, follow its causal path and record a position.
Hazard study
Connect the case to an unverified finding, then optionally inspect the full screening register and authored references.
Controls
Inspect the relevant proposed control, implementation mapping and remaining gap.
Evidence
Distinguish the retained browser result from frozen experiments, inspect the decisive recorded event, then open the complete paired trace when needed.

The nine browser cases cover a currently authorized purchase, unauthorized tool use, excessive spending, recursive delegation, missing approval, replay, approval expiry, conflicting objectives and task-context mismatch. Six secondary cases do not project downstream business effects.

Know which claim each source supports

Approval freshness is not quotation validity. The related recorded context mismatch occurs at demand, while the browser case evaluates a purchase request. The recorded goal-conflict case measures a budget violation, not objective arbitration. APS makes these relationships explicit before presenting the broader source records.

Limits: hazards remain unverified, controls proposed and real human approvals absent from the supplied study records. Passing a check does not establish that the quotation, stock record or demand is correct. There is no real purchase or payment, safety certification, production-protection claim, or established user-learning result.

Where to go next

For a first visit, open APS and follow the approval case. For deeper work, choose the source appropriate to your question: