Andon Labs' Pion: persistent AI agents that run a whole business
Research preview, waitlist only: its agents get banking, a phone, email and a browser to operate a company end to end — plus Tencent's open AuK speech model.
Copy markdown
What Pion actually hands an agent
Launched Sept 14 as a research preview, Pion gives long-running agents real operational tools — banking, a phone line, email, browser automation, a terminal and secure compute — enough to run a company's day-to-day with no human in the loop. It's a 'hand the whole business over' harness, not another chatbot wrapper.
A manager agent watching the workers
Pion runs a two-tier setup: a managing agent (reported as Andonos) oversees persistent business agents that handle daily operations, with monitoring built in to flag agents that go off-mission or try to hoard resources. The pattern — a supervisor over long-lived workers, with guardrails — is where serious agent deployments are converging.
The track record is honest, and thin
Andon Labs has run this for real: a vending machine turned profitable in late 2025, but its Andon Market shop and Andon Cafe, both opened April 2026, still aren't. On the company's Vending-Bench 2, results reportedly improve about $822/month in net worth per model generation. The capability curve is real; the businesses aren't printing money yet.
How to actually get in
Access is waitlist-gated and granted gradually — you sign up at andonlabs.com/pion, and it's aimed at founders and researchers testing autonomous agents. No public pricing. You can't build on it today, but if you have a business (or an idea) to hand off, the queue is open now.
Elsewhere: Tencent's AuK is one open model for all things speech
Tencent open-sourced AuK, a 1.5B speech model under MIT with weights on Hugging Face, that does zero-shot voice cloning, speech and lyric editing, emotion/accent/timbre changes, denoising and speaker separation — all through one natural-language interface. The distilled AuK-Flash runs a fixed 4-step inference (~4.5x faster). Build a whole voice pipeline with one local dependency instead of five.