The fragmentation of AI accelerators has created a prohibitive engineering bottleneck. Supporting a single model across diverse hardware requires weeks of manual porting — an unmanageable N×M combinatorial explosion.
We present FlagOS, an open-source system software stack that consolidates the heterogeneous AI landscape. Core components include a unified operator library (FlagGems, 407+ ops, 25+ accelerators), a multi-target compiler (FlagTree), and a non-invasive plugin system for framework adaptation (FlagScale).
This talk centers on the symbiotic evolution of infrastructure and agents:
FlagOS powers agents. Autonomous agents demand 24/7, high-concurrency inference. FlagOS provides a stable, unified compute abstraction that masks hardware-specific complexities, ensuring reliable inference services across diverse accelerators.
Agents power FlagOS. Through FlagOS Skills — agent-executable workflow packages — we shift from "infrastructure-as-code" to "infrastructure-as-agent." AI coding agents now autonomously perform our most complex infrastructure tasks: 13-step model backporting with E2E precision validation, and MCP-driven kernel generation (99% correctness on 110-operator benchmarks).
The infrastructure is no longer just agent-ready; it is agent-built and continuously agent-improved.
BAAI is committed to global open collaboration. We have contributed PanEval to the Eclipse Foundation as a vendor-neutral AI evaluation platform.
Speaker Slides 
Photo Album 