Vercel Labs open-sourced fx, a coding agent harness and CLI that was, until now, an internal tool. What's unusual isn't the feature set — it's what fx chose to optimize instead of features. Its two headline numbers are a 10-microsecond cold start and a 6.3MiB binary, and the entire design follows from taking those two numbers seriously.
Built to disappear, not to impress
fx is written in Zig — a genuinely unusual choice for an agent harness, where TypeScript, Python, and Rust dominate — specifically because Zig compiles to a single static binary with no runtime to install and no garbage collector pausing execution mid-task. The binary is 6.3MiB, uses single-digit megabytes of memory at baseline, and does "no unnecessary work or I/O before accepting input." Minimalism runs through the whole stack: a small system prompt, a small core tool set, features added through skills, plugins, and MCP rather than built in. That's the opposite instinct from oh-my-pi, the community harness this blog covered last week — 31 built-in tools, persistent Python/JS interpreter cells, an advisor role running a second model as standing critic, a whole daily-driver feature list. Both are legitimate design philosophies for the same category of tool; fx's is that the core should be small enough to embed anywhere, and everything else is someone else's plugin to write.
The pitch is infrastructure, not a product
fx's own framing is explicit about who it's for: "optimal for model benchmarking, sandboxing, evals, and gyms." That's the vocabulary of the RL research infrastructure this month's Prime Intellect coverage was built on, where the harness itself is the variable being tested, spun up and torn down thousands of times across sandboxed nodes. A cold start measured in microseconds and a footprint measured in single-digit megabytes matter enormously at that scale in a way they don't for a human typing into a terminal once. fx ask --json gives scriptable structured output for exactly this use case; fx acp speaks the Agent Client Protocol to connect to editors and other clients; a WebAssembly build runs the entire harness inside a browser tab at fx.sh/try — three different embedding surfaces for three different audiences, none of them "install this and use it as your primary coding assistant."
Privacy as an engineering constraint, not a policy promise
The privacy claims are worth reading closely because they're structural rather than aspirational: no product telemetry, sessions and usage stay local, and — the specific, checkable line — no source code or prompts are shared with any endpoint other than the inference provider itself. Pair local inference with auto-updates disabled and fx becomes, in its own words, fully hermetic: nothing leaves the machine at all. That's a stronger and more specific claim than the usual "we respect your privacy" boilerplate, and it's the kind of design decision that only makes sense for a tool meant to run inside a sandbox or an eval harness where data exfiltration isn't a policy violation to prevent after the fact — it's an architecture to make impossible from the start, the same instinct behind OpenAI's own post-breach network isolation work this week, applied from the opposite direction: not containing an agent that might misbehave, but building a tool with nowhere to leak to in the first place.
What to expect next
- Watch for independent cold-start and footprint verification. Both headline numbers are trivially reproducible by anyone with the binary, which is a genuinely low-friction fact-check to run.
- Watch whether fx gets adopted as the substrate underneath other labs' evals. Its own stated use case — benchmarking, sandboxing, gyms — is exactly the layer Prime Intellect's 153-run experiment needed and built custom tooling for; a shared, minimal, embeddable harness could become plumbing other research reuses rather than reinvents.
- Watch the skills/plugin/MCP ecosystem that grows around it. A deliberately minimal core only stays useful if the extension layer fills in fast — the same test oh-my-pi's plugin architecture already passed from the maximalist side.
References: fx.sh · fx.sh/try — WebAssembly demo · related coverage: The Harness With 23,600 Stars Wasn't Built to Win a Benchmark · Prime Intellect Ran 153 Autonomous AI Research Runs · OpenAI Paused Its Biggest Training Run Because One Model Might Be Too Good at Cyber · Frontier Arcade: trends & predictions