AI agent tooling: Monty and Beads

Two projects worth keeping an eye on.

pydantic/monty

A minimal Python interpreter written in Rust, built to run LLM-generated code safely without a container sandbox. Startup is measured in microseconds (~0.06ms), and isolation is strict — filesystem, env vars, and network access are all blocked by default and only exposed through explicit host callbacks you control. It supports a deliberate subset of Python (no classes yet, no third-party libraries), which is the point: the goal is a safe execution surface for agent code, not a general runtime. The snapshot/resume capability (dump()/load()) lets you serialize interpreter state mid-flight and resume it later, even across process boundaries. Still experimental, but it will power code-mode in PydanticAI — where the LLM writes Python that orchestrates tool calls rather than invoking them one by one via the usual tool-calling API.

gastownhall/beads

A dependency-aware graph issue tracker for coding agents, backed by Dolt. The premise is that markdown task files are a poor fit for agents working on long-horizon tasks — they go stale, they have no dependency semantics, and concurrent agents step on each other. Beads replaces them with a version-controlled SQL graph: tasks have typed relationships (blocks, relates_to, supersedes), hash-based IDs prevent merge collisions across branches, and a compaction step summarizes closed tasks to avoid context window bloat. The embedded mode runs Dolt in-process with no external server; server mode adds concurrent multi-writer support. Already at v1.0.0 with 20k stars — this has clearly hit a nerve.