Blog

# Customer support and operations agents

Use cases / September 22, 2026 / Krishna Bhatnagar

Refunds, exchanges, cancellations, account updates. Customers ask in a thousand ways; the jobs underneath are a short list.

A support agent at an airline, a retailer or a telecom spends its day on a short list of jobs. Where is my order. I want a refund. Cancel my booking. Change the address on my account. Every customer words it differently; the job underneath doesn't change.

That makes support the clearest fit for AgentCompile, and it's the work τ-bench tests: an agent with tools, a policy to follow and a simulated customer on the other side. It's where our results come from.

## What changes

Known jobs run compiled. On τ-bench retail, with simulated customers, that meant 58% fewer agent calls, with the same answers. Anything new, unclear or unusual still goes to your agent, unchanged, with the whole conversation so far.

## What stays protected

- A refund or a cancellation never happens without the customer's explicit yes.
- If AgentCompile has any problem, the call goes straight to your model.
- A job goes live only after it gets your past conversations right.

## One job, a thousand wordings

Take a single job: telling a customer where their order is. In a day, a support agent sees it as "where's my stuff", "order 1042 hasn't arrived", "tracking says delivered but it isn't here" and "can you check on my package". Four messages, four moods, one job: find the customer, find the order, read its status, reply.

Today your agent's model reasons through that job every single time. It reads the policy, reads the conversation, decides which tool to call, calls it, reads the result and writes back. The thousandth time costs as much thought as the one before it.

Without AgentCompile:

- Every message goes through your agent's model.
- The same lookups are reasoned out again for every order.
- A refund is decided from scratch each time it comes up.

With AgentCompile:

- Known jobs run compiled, without calling your agent's model.
- New details (a new customer, a new order) are handled by the compiled job.
- Anything new, unclear or unusual goes to your agent, unchanged.

## A refund, walked through

1. The customer asks for a refund on an order.
2. The job is known, so it runs compiled: find the customer, find the order, check it against your policy.
3. Before anything changes, the customer is asked to confirm. Nothing irreversible happens without their explicit yes.
4. The customer says yes, and the refund goes through. If they say something unexpected instead, the conversation goes to your agent with everything so far.

## Where the numbers come from

τ-bench retail is a public benchmark built on exactly this work: an agent with tools, a policy to follow and a simulated customer. On it, 34% of held-out conversations finished end to end with no agent call, and compiled writes missed the right answer less often than the agent alone: 11.6% against 15.6%.

## Getting started

1. Show us your agent's logs. We show you which jobs repeat, and how often.
2. We prove each job on your history. You see how each one did before anything goes live.
3. Wrap your model client. Known jobs run compiled; everything else goes to your agent as usual.

See every number, with its source. [Read the results](/blog/58-percent-fewer-agent-calls)