Coding agents change how software is written. Maxwell is pitched at how it runs: it watches a live program, rewrites the binary underneath, and leaves the source humans and agents read exactly as it was.
In brief
- doubleAI (19 October 2026) has opened early access to Maxwell, which it calls the first Code Execution Model (CEM) — AI that acts “on behalf of the hardware” rather than writing source for a programmer.
- For teams burning GPU and CPU cycles on pipelines and kernels, the company claims Maxwell has made data pipelines over 38× faster and AI inference kernels over 4.5× faster on real-world workloads, while formally proving each change leaves behaviour unchanged.
- What’s still unproven outside the lab: those speed-ups are doubleAI’s own, and access is early only at heymaxwell.ai.
Agents for the machine, not the developer
Co-founder and CEO Prof. Amnon Shashua put the distinction bluntly in the Business Wire release: “AI has changed how software is written, but not how it runs… they are the first AI that works for the machine, not the programmer.”
That pitch lands in a compute-scarce market. If idle FLOPs can be unlocked without touching application code, the buyer is every team that already trusts a compiler pass more than a chat window. The catch doubleAI itself highlights is trust: runtime rewrites get no human review, so Maxwell is sold with “formal and mathematical” proofs that behaviour is unchanged.
| Claim (doubleAI) | Figure |
|---|---|
| Category | Code Execution Model (CEM) |
| Access | Early access at heymaxwell.ai |
| Data pipelines | >38× faster (company) |
| Inference kernels | >4.5× faster (company) |
| Source code | Unchanged; binary rewritten |
Maxwell, the company says, runs on CPUs and GPUs, keeps improving from each proven faster binary, and never alters the source coding agents still write. That framing positions CEMs as a complement to coding agents, not a replacement chat bot.
Background funding, not today’s news
The same release notes that doubleAI raised a $200 million Series A co-led by Lightspeed and Bessemer, with Dell Technologies Capital, Greenfield, NVIDIA, Pitango and BRM also named. That financing is background context for who can afford to ship a CEM; it is not a new round dated today.
Early access at heymaxwell.ai is the only distribution channel named in the release. doubleAI says Maxwell continuously inspects and improves the binary under the code as it runs, fitting it to the exact hardware, data and usage it encounters — and that every rewrite is formally proven to leave program behaviour unchanged.
That verification story is what separates the pitch from a conventional JIT or profile-guided optimiser sold with marketing numbers. Until customers publish independent before-and-after benches, the 38× pipeline and 4.5× kernel claims remain company figures attached to “real-world workloads.”
Our read
If Maxwell’s proofs hold under hostile workloads, CEMs become a new layer in the stack between kernels and silicon — complementary to coding agents, not a rival chat bot. Until independent benches appear, treat the 38× / 4.5× lines as marketing ceilings. The open question is whether early-access users will publish before-and-after numbers that survive outside doubleAI’s demos.
Sources
Related stories
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
