Blog

# Fewer agent calls on every kind of agent we measured

Results / October 8, 2026 / Krishna Bhatnagar

Browser, customer support, retail, voice and coding agents, each on a public benchmark with simulated users. What changed, and how to read it.

We measured AgentCompile on five kinds of agent, each on a public benchmark with simulated users. In every one, the agent made fewer agent calls with AgentCompile than without it.

- Browser agents: 64% fewer agent calls.
- Customer support agents: 56% fewer agent calls.
- Retail agents: 55% fewer agent calls.
- Voice agents: 45% fewer agent calls, and 2x faster.
- Coding agents: 30% fewer agent calls.

## Why the numbers differ

How much an agent saves depends on how much of its work repeats. A browser agent that goes through the same pages for different customers repeats a lot. A coding agent meets more work it hasn't seen before, and anything new goes to its model, unchanged.

## How to read them

- These are public benchmarks with simulated users, not customer data.
- Fewer agent calls means fewer calls to your agent's model. It is not a cost figure.
- What you save depends on your own agent's traffic. A pilot measures that.

Every number, with its source. [See the results](/results)