400+ tool calls per turn: AIOps platform uses Monty's 1.2ms Rust sandbox to keep MCP output out of context
samuelcolvin · x · 2026-10-29
Monty, Pydantic's minimal secure Python sandbox written in Rust for running LLM-generated code, is proving itself in production. It ships as OSS Monty (MIT-licensed Python 3.14 sandbox package) and Full Monty (commercial WebSocket service).
Latency is the headline: fresh sandbox + agent run costs 1.2ms for OSS Monty and 7ms for Full Monty, vs 200ms for WASI/wasmtime, 900ms for local Docker, 1900ms for sandboxing services, and 2700ms for Pyodide in Deno — roughly two orders of magnitude below container-based options.
Real-world validation: an AIOps platform integrating thousands of third-party MCP servers reports tasks that make 400+ tool calls per turn. With Monty's code mode, agents write sandboxed code to transform and filter tool outputs, preventing that flood from hitting the model's context directly.
More from coding & agent
- Build-A-Chip-Bench: end-to-end chip design benchmark, top agent scores just 16.6% — teortaxesTex · 2026-10-29
- Separate Discovery from Scoping: Grill the Agent Before It Writes Code — rseroter · 2026-10-29
- Autocomplete.dev brings Tab-style next-message suggestions to chatbots, 20k users already — josh_bickett · 2026-10-29
- Coinbase Agent trading hits all-time highs, with Grok Bot taking 67% of volume — MurrLincoln · 2026-10-29
- REA Turns App Recordings into RL Environments; 'Wave of Sloppy Envs' Incoming — teortaxesTex · 2026-10-29
- 86M-param decoder-only transformer plays chess on Lichess, full notebook public — andreisavu · 2026-10-29