A 6MB coding agent, and why fleets need one

Nano Workforce turns software delivery into durable graphs: plan, implement, review, test, merge. The interesting graphs fan out — ten slices of an epic, each with its own coding agent, each looping through review until it converges. At that point the question stops being how good is one agent and becomes how many can this machine hold?

The footprint is the ceiling

Node-based agent CLIs sit at 150–660MB of resident memory each. Multiply by a fleet and the laptop is full before the work is. Most of that memory does nothing: an agent spends its life waiting on a model or a test run.

nano-coder

nano-coder is a coding agent written in Rust that runs in about ~6MB. It is a complete terminal agent — streaming REPL, tool calling, task plans that survive compaction, OS sandboxing, AGENTS.md and skills — and it also runs headless over ACP (JSON-RPC 2.0 on stdio), which is how Workforce drives it as a durable job worker.

Because the engine owns durability, the worker doesn't have to: if a nano-coder process dies, the Nano instance that dispatched it is still at the same step, and the next worker picks the job up. Small workers plus a durable orchestrator is the whole design.

Try it

npm install -g @nanobpm/nano-coder

Prebuilt binaries ship for macOS and Linux (x64, arm64); or cargo install nano-coder. Source and docs are on GitHub.