Maze: A scheduler that treats LLM agents like batch jobs

What it does Maze is a distributed runtime for LLM agent workflows. You define tasks with Python decorators, wire them into static or dynamic graphs, and submit them to a head-worker cluster that handles CPU/GPU allocation, retries, timeouts, and cancellation. It also supports ReAct-style agents where the LLM decision loop itself is a task graph, plus plain application specs via YAML for non-agent jobs.

The interesting bit The project treats agent execution as infrastructure rather than inference plumbing. It records durable run snapshots—lifecycle events, structured errors, placement, artifacts—so you can inspect or retry a run without holding a live WebSocket open. For distributed setups without shared storage, it moves workspace files through a content-addressed artifact store (maze://artifacts/sha256/...) so workers can fetch inputs and upload outputs without mounting the same filesystem.

Key highlights

• Task-level resource requests (cpu, cpu_mem, gpu, gpu_mem) prevent contention across parallel workflows
• Dynamic workflows can append tasks at runtime; ReAct workflows keep LLM decisions and tool calls in the same event history
• LangGraph workflows can reportedly migrate to Maze without logic changes to gain task-level parallelism
• A web-based Playground offers drag-and-drop workflow construction, workspace file management, and natural-language task generation
• Content-addressed artifact store enables distributed execution without shared filesystem assumptions

Caveats

• The README notes that Ray workers must explicitly register as Maze workers before the scheduler will place tasks on them; joining Ray alone is insufficient
• ReAct skills rely on a specific SKILL.md directory convention, and the compact catalog exposure means the LLM may need to call back for full skill files
• Several features (sandbox, Playground task generation, artifact store) are mentioned but not deeply documented in the README; the docs site may cover them

Verdict Worth a look if you’re running agent workflows that outgrow a single machine and need observability, retries, and resource isolation that feels more like a job scheduler than a chat loop. Probably overkill if your agents run comfortably on one GPU and you don’t need durable run history or cluster-wide queueing.