Support investigations with a disposable AI agent
A support agent looks up the payment, ticket or thread behind a customer question and writes the cause and the next step for your team. Ephemeral starts a fresh microVM for each question, gives it one read-only connection with only the methods you allow, reserves the budget before it starts and delivers the result by signed callback or API, then destroys the machine.
A payment investigation took 12 seconds and cost $0.007651, measured on October 7, 2026 with openai/gpt-6-luna (3 model calls, 5,101 tokens).
What the run is allowed
Everything a run can touch is listed in its request. This is the request of the sample AI payment failure investigator: why did this payment fail?.
| Files | playbook.md (read-only) |
|---|---|
| Connections | payments: GET only; credentials stay outside the machine |
| Web access | ephemer.al, search off |
| Shell | Off |
| Budget | $0.04 per run, reserved before it starts |
| Timeout | 90 seconds |
How to run it
Start a run with the API or the CLI. The result arrives by signed callback or when you read the run.
POST /v1/runs
{
"workflow": "investigate-payment",
"input": {
"payment_id": "pay_123"
},
"max_cost_micros": 40000
}Samples
Each sample is a complete request you can read, replay and copy.
- AI payment failure investigator: why did this payment fail?An agent reads a payment, its attempts and the processor status page, then explains the cause and the next step for your support team.
- AI support ticket triage: classify, prioritize and routeAn agent reads a ticket and your routing rules, then returns the category, priority, owning team and a suggested first reply.
- Summarize a support thread with an LLM, with zero data retentionAn agent turns a long support conversation into the issue, what was tried, the customer's mood and the open question, and keeps nothing afterwards.
- Draft replies to refund requests with AI, never send themAn agent checks a refund request against your policy and drafts a reply for a human to approve. It cannot issue the refund.
- Prompt injection example: what happens when a ticket tries to hijack the agentA hostile ticket tells the agent to send its API key to an attacker and issue a refund. See why it cannot, and what the timeline shows.
Recipe
Recipe: investigate a failed payment walks through the same job step by step.
Related terms
- Credential broker: Ephemeral holds a connection's credentials and adds them to each request, so the token never enters the machine.
- Run budget: The most a run may spend, reserved up front and enforced while it runs.
- MicroVM: The small virtual machine each run gets fresh and that is destroyed when the run ends.
Questions
Can the agent refund a payment or change a ticket?
Not unless you let it. The connection lists the methods it may call, and the support samples allow only GET, so a refund request is refused by the connection even if the model is tricked.
Does the agent ever see our API key?
No. Ephemeral holds the credentials and adds them to each request outside the machine, so the key is never in the sandbox for a prompt to extract.
What if a customer message tries to hijack the agent?
The prompt-injection sample runs exactly that case: a ticket that asks the agent to send its key to an attacker and issue a refund. It has no key to send, a read-only connection and one allowed domain, so neither action can succeed.
Prices are on the pricing page.