Building an agentic world.
Tokens and agents could power a new industrial era. Turning model capability into useful work takes systems that coordinate action, check results, and improve from experience. That is what we build and test.
- Coordinate: Give agents clear tasks, shared context, and recoverable handoffs across tools and teams.
- Verify: Check results against the task, with evidence people can inspect and boundaries they control.
- Improve: Use outcomes, failures, cost, and human effort to improve the next run.
- Small models for agent work: fine-tuning small open models on the tasks agents actually do.
- Verified math: agent teams that solve math problems and verify their answers.
- Efficient inference: serving open models on our own GPUs at the lowest cost per token.
- One gateway for agent fleets: a single access point to many models, with usage, cost, and limits per team.
- zagent: An unofficial, GLM-native terminal coding agent. Early release.
