The travel agent for GPU jobs

What it does

SkyPilot is a job launcher and scheduler that treats infrastructure as a menu. You write a task once—resource requirements, setup scripts, run commands—and it finds the cheapest available GPUs across your clusters, clouds, or Slurm grids, provisions them, syncs your code, and streams logs back. When the job finishes, it can autostop idle resources so your finance team doesn’t send angry Slack messages.

The interesting bit

The project leans hard into “BYOC” (bring your own cloud): everything launches inside your accounts and VPCs, so you’re not handing data to a managed platform. The scheduler does auto-failover between providers—if AWS is out of A100s, it quietly tries Lambda or RunPod without you rewriting anything. There’s even a “Skill” for Claude Code and Codex, so your agent can provision its own GPUs. The 2026 research-agent demo—reading arXiv papers, then fusing llama.cpp kernels for $29—is either impressive or slightly unnerving, depending on your worldview.

Key highlights

• Unified YAML/Python API targets Kubernetes, Slurm, and 20+ clouds (AWS, GCP, CoreWeave, Vast.ai, Crusoe, etc.)
• Gang scheduling, multi-cluster pooling, and binpacking for shared GPU fleets
• Auto-failover and cost-based provisioning across providers
• Agent integration: LLM coding agents can directly launch and manage jobs
• Shopify and H Company case studies suggest production traction

Caveats

• The README is heavy on quickstart snippets and light on architecture details; you’ll need to dig into docs for multi-team deployment patterns
• GPU quota and cloud account setup remain your problem—SkyPilot finds capacity, but can’t conjure it

Verdict

Worth a look if your team is drowning in cloud consoles and Slurm queues. Skip it if you’re already happy with a single vendor’s managed notebooks or your workloads never leave one Kubernetes cluster.