Show HN: a1 - determinism-maxing JIT compiler for AI agents (github.com)

a1 is an open-source “agent compiler” that compiles AI agents into executable code ahead-of-time (AOT) or just-in-time (JIT), aimed at making multi-tool agents safer, faster, and more deterministic. Instead of running agents as a static while-loop, a1 generates code per input, isolates LLM contexts to minimize exposure to untrusted data, and applies static checks and aggressive parallelism to speed up execution. It’s installable via pip and treats LLMs as first-class tools alongside user-defined functions (e.g., Pydantic schemas + async tools), with simple APIs for AOT compile.execute(...) and jit(...). For the AI/ML community this matters because it tackles three persistent pain points: nondeterminism, safety, and performance. a1 optimizes for determinism with a swappable cost function and a Generate/Cost/Verify workflow that lets the compiler explore execution plans under engineered constraints (you can even preserve loop semantics while letting the compiler optimize it). Features include LangChain import, OpenTelemetry observability, tool instantiation from MCP/OpenAPI/FastAPI, RAG over SQL or fsspec (S3/GCS/local), and zero vendor lock-in for LLMs or secure execution clouds. The project emphasizes auditability, static checking, and composable context engineering—making agent deployments more reproducible and production-ready while enabling research on better code-generation, costing, and verification strategies.