How does Decagon issue refunds?
Decagon defines an AOP as ‘a step-by-step set of natural language instructions backed by code’. Each AOP has a name and entry conditions that tell the agent when to use it, and content blocks add conditional logic. Decagon names refunds and account updates as the kind of sensitive operations AOPs are designed to handle with guardrails.
The money moves through tools, which Decagon describes as ‘custom logic like API calls, data lookups, or calculations that the AI agent can run in real time’. Metadata such as location, role or product tier lets an AOP adapt to the customer.
Decagon’s public pages describe this at the product level. Configuration details, such as how a refund limit or an approval step is set inside an AOP, aren’t in public documentation, so check them in your own workspace.
Which guardrails does Decagon describe?
Decagon says technical teams add guardrails, integrations and advanced logic to AOPs, while CX teams write the procedures. At runtime it describes two further layers.
| Control | Where it lives | What Decagon says |
|---|---|---|
| AOP logic | Decagon | Natural-language steps backed by code, with conditional blocks |
| Supervisor model | Decagon | Monitors live conversations for safety issues |
| Escalation intents | Decagon | Route high-risk topics away from the agent |
| Refund tool limits | Your API | Not described by Decagon. Caps, eligibility and idempotency checks sit in the system the tool calls |
What do Decagon’s simulations cover?
Simulations run AI-generated conversations between your agent and a mock persona across channels. Decagon groups what they check into three areas: AOP selection and execution, knowledge base retrieval, and guardrails and brand tone, which covers sensitive topics, escalation to humans and staying on-brand with frustrated users.
Personas and intents can be generated from real failure modes in uploaded historical transcripts. Named traits include emotional states such as anger, stress and frustration, and voice tests add accents, speaking speed, background noise and interruptions. You can configure metadata, tool outputs and conversation length. Each run returns pass or fail with a trace of the decision path.
Decagon says simulations integrate with CI/CD so every agent version is tested before it merges, and that its Duet assistant maintains a golden test set generated from AOPs and production conversations.
What do the simulations leave out?
Decagon’s simulation posts describe personas by mood and channel, not by intent to extract money. None of the pages reviewed mentions adversarial personas. A calm customer who invents a prior approval or quotes a fake policy is a different test from an angry one.
The posts don’t say whether tools call real systems during simulations. They describe configuring tool outputs, which suggests simulated responses. Either way, your refund API’s own limits need tests of their own.
What to test before giving a Decagon agent refund authority
Illustrative cases to add to your simulation set:
- A polite customer who says an earlier agent promised the refund (fake prior approval).
- A request that starts inside one AOP and drifts into a refund over several turns (multi-turn escalation).
- A “damaged on arrival” claim on an order with proof of delivery and no photos (refund fraud patterns).
- Pasted text claiming a policy exception (indirect prompt injection).
- Questions that probe for the refund threshold before the ask (information extraction).
- Two refunds on one order, checked against the tool output and your API.
Questions
What is an AOP in Decagon?
An Agent Operating Procedure is a step-by-step set of natural-language instructions backed by code. It has entry conditions that tell the agent when to use it, and it calls tools such as API calls, data lookups and calculations to act.
Do Decagon simulations include adversarial customers?
Decagon’s published descriptions mention personas with emotional states such as anger, stress and frustration, generated partly from real failure modes in historical transcripts. They don’t mention adversarial personas, such as customers inventing approvals or quoting fake policy, so add those cases yourself.
Can Decagon simulations run in CI?
Yes. Decagon says simulations integrate with CI/CD pipelines so every agent version is tested before it merges, and that outdated tests are detected and updated or removed automatically.
Decagon is a trademark of its owner. This guide is independent and not affiliated with or endorsed by Decagon.
Free checklist
Get the guardrail checklist for agents that move money.
A one-page PDF of the controls and the test for each. Download it right away, and we email you the link.
Sources
- Decagon: Agent Operating Procedures
- Decagon: Agent architecture
- Decagon blog: Decagon Simulations
- Decagon blog: The next generation of simulations
- Decagon: Testing and QA