Zeus – Claude Code for AI voice agents

ThunderPhone, a company that runs AI agents to answer and place phone calls for businesses, has launched version 3 of its platform along with Zeus, an AI agent whose job is to build and tune those phone agents. CEO Alex Kolchinski's launch post explains the "Claude Code for voice agents" label by analogy. Anthropic's coding agent can check its own work by running tests, but general coding agents cannot, out of the box, run real or simulated calls through a voice agent, so they have no reliable way to tell whether it works. Zeus is given that check. It can create an agent from a conversation about the business or from imported call recordings, scripts and agents built on other platforms, run simulated calls in text and audio, flag problems on live calls, and propose A/B tests of fixes. Customers decide how much autonomy it has, from asking before every change to acting on its own.

ThunderPhone wants an agent for this job because its own system has so many settings. The usual design for phone AI is a three-step pipeline: a transcription model turns the caller's speech into text, a language model writes a reply, and a text-to-speech model speaks it. The post argues that this design loses information when the audio is reduced to one best-guess transcript, and forces a choice between fast language models that make mistakes and accurate ones that take seconds to answer. ThunderPhone instead runs several transcription models at once, often sends the audio itself to a language model, pairs a fast and a slower language model and picks between them turn by turn, and adds models for noise, separating speakers and detecting when someone has finished talking. Until now, self-serve customers chose among three preset configurations tuned for cost, speed or intelligence, while enterprise customers got hand-tuned ones. Version 3 retires the presets, and Zeus tunes a single architecture to each customer's goal, such as more conversions, fewer hang-ups or a hard cap on cost per minute.

The performance claims are ThunderPhone's own. The post says its system makes significantly fewer mistakes than any three-step or speech-to-speech system the company has seen, and that it holds the top score on the Big Bench Audio benchmark, 99.4%. It publishes no comparison data and says more benchmarks, including public versions of its internal evaluations, are still to come. Zeus is free for now and will be priced later, simulated calls are billed like other test calls, and the company describes Zeus as an early product. The approach of an agent that edits a voice agent and tests the change on simulated calls is not unique to ThunderPhone: ElevenLabs earlier introduced a similar assistant, Architect, inside its own agent platform.