Agent Pattern Labs builds open infrastructure for AI agents that need to run outside demos: smaller runtimes, durable state, verifiable work, model routing, and practical tools for self-owned infrastructure.
Our goal is to make agent systems easier to inspect, cheaper to operate, and more portable across the tools developers already use.
- Agent runtimes and orchestration: local execution kernels, durable agent state, leased tools, browser/session management, and multi-agent workflows.
- Verification and evals: proof challenges, traces, and contracts that make autonomous work easier to check.
- Model and tool routing: infrastructure for fallback, fan-out, guardrails, ledgers, and cross-tool compatibility.
- VPS-friendly systems: packages and reference tooling for people running their own agent infrastructure.
- Applied workflows: experiments and products that use these primitives in real developer and operator tasks.
- ray: shrink AI to run on cheap VPS infrastructure.
- sparse-kernel: local multi-agent runtime with durable SQLite state, leased tools, artifacts, and agent specs.
- state-trace: agent state infrastructure for working memory and durable MCP orchestration.
- agent-proof: action-bound proof challenges for verifying fresh autonomous AI-agent work.
- iso: isomorphic agent tooling for workflows across Cursor, Claude Code, Codex, and OpenCode.
- geometra: computed geometry and browserless automation tooling.
- Local-first when possible.
- Operator-owned infrastructure over hidden hosted state.
- Durable traces over opaque sessions.
- Small, composable packages over monoliths.
- Proof, measurement, and reproducibility over unsupported claims.
Most projects here are infrastructure pieces, research prototypes, or applied experiments. The common thread is practical agent systems that can be understood, operated, and improved by the people running them.