Many minds, one answer
The lion, the goat, the dragon and the snake were on the logo before they were an argument. A panel of models answers, a judge names where they agree and where they do not, and a synthesiser writes the result — but only when the question is hard enough to be worth it. Fusion itself is not unique; wiring it into the agent loop behind a cost-aware router, and measuring it, is.
Three bets, and where each one stands
Evolution with a fitness signal
Other agents learn by appending what happened. Chimera keeps a learned change only when a verified result proves it helped — gated on the real working-tree diff and an honest A/B, never on what the model says about itself. Whether that accumulation makes it better at new tasks is measured, and the answer so far is on the evidence page.
Security as architecture
Prompt injection is widely considered unpatchable, and most agents mitigate it in the application layer or declare it out of scope. There is a real defence layer here — taint tracking by verbatim reference, not true dataflow (a model that paraphrases tainted text launders it), control tokens stripped from untrusted content, dangerous tools narrowed for the rest of a tainted run, side-effecting retries protected. It is opt-in: armed with --taint, off by default. The security page in the documentation states where each layer stops, and names what still gets through.
Benchmarks published, including the losses
Every published figure comes with a confidence interval, the runs that went nowhere are published unchanged, and a claim that stops surviving replication is withdrawn. Nothing on this site is typed by hand; it is read from the snapshot the product stamps at each release, and the intervals are on the evidence page.
What is actually measured
Two products
Terminal agent
The harness: a right hand in your terminal. It plans, edits, runs the check, and reverts itself when the check fails.
Read more →Desktop app
The app: conversation that edits code, a board that dispatches to your agents, and every setting the agent has — on Windows, macOS and Linux.
Read more →
A skill library, with provenance
Short cards that tell the agent how to approach a recurring kind of task — data, never code; nothing there is executed. There are few of them, and each has a name behind it: reviewed by a maintainer before it merged, served with a hash of the exact bytes on the page, and read in your language — a translation that has drifted from its source falls back to English rather than paraphrase a sentence that no longer exists.
It is alpha, and it says so
Solid and heavily tested, not yet battle-hardened in production. The installers are not signed by a certificate authority, so Windows and macOS warn on first run. That warning sits above the buttons on the download page; the alpha note is at the end of it.