Agent Operating Procedures: Why Every Workflow Needs a Decision Trace

The enterprise trust problem

Enterprise customers will not deploy AI agents that make decisions they cannot audit. Supply chain operations involve financial commitments (booking freight, accepting appointments, disputing invoices), regulatory compliance (customs documentation, chain-of-custody records), and customer relationships (delivery commitments, proactive notifications). Every decision an agent makes in these domains must be traceable: what data was used, what logic was applied, what alternatives were considered, and what outcome was produced.

The AOP framework

Every workflow deployed on Loft is defined as an Agent Operating Procedure (AOP). An AOP is not just a workflow definition. It is a complete specification that includes the trigger conditions, the decision logic, the actions taken, the data sources consulted, the confidence thresholds applied, and the escalation rules for cases that exceed the agent's authority.

The decision trace

Every time an AOP executes, Loft records a decision trace: a timestamped log of every input, every decision point, every action, and every outcome. When Tracy decides to escalate a late shipment to the shipper's operations team rather than sending a second follow-up to the carrier, the decision trace records why: the carrier's reliability score on this lane is below threshold, the shipment is within 4 hours of the delivery commitment, and the shipper's configured escalation rules specify human involvement at this severity level.

The lifecycle: Draft, Tested, Deployed

AOPs follow a formal lifecycle. A new workflow starts as a Draft: the logic is defined but not executing in production. It moves to Tested when it has been validated against historical scenarios (by the Test Agent in Sophie's pipeline). It moves to Deployed when it is actively executing in production with full decision tracing. At any point, an AOP can be paused, modified, and re-tested without affecting other workflows.

Why this matters

Decision tracing is the difference between AI that enterprises will adopt and AI that they will evaluate forever. When a VP of Supply Chain asks 'Why did your system rebook my ocean freight?', the answer is not 'the AI decided.' The answer is a specific decision trace showing what data was used, what logic was applied, and what alternatives were considered. This is what makes FourKites agents enterprise-grade.