Rust is being readied for AI agents: Meta wrote twice as much Rust code in a year as in the previous 9

That's the argument Niko Matsakis makes in an Oct 14 write-up. He leads the Symposium project at the Rust Innovation Lab, and he borrows his point from Dijkstra in 1970: testing does not prove the absence of bugs.

The discussion goes beyond Rust. When an agent writes the code, the volume grows faster than a human can read it, and the guarantee no longer comes from review but from what a machine checks.

The growth figure. Meta engineers wrote twice as much of their own Rust code this year as in the previous nine years combined. The figure comes from a post by David Tolnay.

On Jan 2, Greg Brockman called Rust the ideal language for agents: if it compiles, it's almost certainly correct. Matsakis splits the work into three parts: figuring out what the code should do, writing it, and making it fast. Agents are strong at the second and third, the intent stays with humans, and authors of cryptographic algorithms are already tuning performance with agents.

What to install today. The post names a working setup for network services: with a profiler like dial9, agents do a decent job finding async pitfalls, long polls and blocking calls. The profiler installs with `cargo install --locked dial9 --features cli`, Cargo.toml gets `dial9 = { version = "0.5", features = ["cpu-profiling"] }`, and .cargo/config.toml gets the rustflags `--cfg tokio_unstable` and `-C force-frame-pointers=yes`. Prebuilt binaries are available for Linux, macOS and Windows, so building from source isn't required.

The same author's second tool takes two commands: `cargo binstall symposium`, then `cargo agents init`. It finds and wires up skills, hooks and MCP servers based on your crate's dependencies. The Rust Foundation approved Symposium as the second Rust Innovation Lab project on Apr 21 and describes it as early and fast-moving.

On top of this, Matsakis proposes a goal for Rust in the agent era: no unsafe code that hasn't been proven correct. Unsafe used to be covered by review and tests; now the proposal is a proof that rules out undefined behavior.

There are four tools for this: Verus, Creusot, Aeneas and Kani. Matsakis advises starting with proving that the code doesn't panic, which removes a whole class of bugs. Assertions on the expected result come next.

Taken further, this logic leads to a Rust where a proof of correctness stands where a test stands today.