Skip to main content
We just open sourced a tiny GPT-style cognitive core built in pure Rust.See our repository
Blog

Introduce Ante: Self-Contained Agent That Self-Organize

· 4 min read
Antigma
Introduce Ante cover

Today Anthropic "open-sourced" Claude Code — and it's the perfect day to introduce Ante, the precursor of our mission, a self-contained agent built from first principles.

And why yet ANother TErminal agent? Well, "it's just built different."

Built Different​

Here is a simple benchmark running same 20 parallel containerized tasks. Same models. Same workload. The only thing different is runtime.

Compared with Claude Code, Ante consumes:

  • ~7x less peak memory
  • ~9x less average CPU
  • ~5x less disk I/O

while achieving same accuracy.

We believe agents should be like cells in a living organism — tiny, expendable, massively replicated. Lightweight enough to run by the thousands. Reliable enough that the system self-heals when any one of them fails.

For one agent doing one task, none of this matters. The model is the bottleneck. You could wrap the API in a shell script and it would mostly work.

But we're not building for one agent. We're building for hundreds — thousands — of agents that coordinate, delegate, fail, recover, and keep going. At that scale, every byte per instance, every wasted cycle, every unnecessary dependency gets multiplied a million times.

We call this the cellular-native thesis. Three pillars:

  • Lightweight. Each replica can't cost gigabytes of memory — especially now, when RAM is expensive. Every line of code is a liability. Keep the core tight.
  • Reliable. The return on reliability is non-linear. Below a threshold, the system is unusable. Above it, the game changes. There's a phase transition — you need to be on the right side of it.
  • Closed-loop. Orchestration should be self-repairing — converging to a desired state automatically. Homeostasis in biology. Kubernetes in infrastructure. Individual agents are expendable; the organism persists.

Measure the Harness, Not the Model​

We started with eval first. Benchmarks are the primary way to get real data before getting users. The principle: we would rather burn tokens than burn customer trust.

We use Terminal Bench and Harbor as our primary external benchmark. The team behind Harbor and Terminal Bench are genuine researchers — rigorous, principled, and motivated by getting it right rather than hype.

We're proud of the results — topped Terminal Bench 1.0 and Terminal Bench 2.0.

Most of the magic comes from the model — but the agent harness is the critical conduit between human and AI.

We evaluate the agent and how well it channels the model's power — not the model itself.

We aim to improve the harness, not the prompt.

More in our docs.

Origin​

Ante started as a side project to run open models on multiple heterogeneous commodity hardware with basic agentic capabilities backed by MCP, focusing on inference in embedding space.

Ante started as a side project to run open models on multiple heterogeneous commodity hardware

Then Claude Code showed us — and the industry — how much agent development is still uncharted territory. Full of excitement and unsolved problems, and critical to our mission of self-organizing intelligence.

Since then, we went all in, deep into the token space.

Self-Contained First​

However, we also didn't forget our commitment to privacy and individual empowerment

Sovereign Compute and Network State, Second Amendment of the AI Era

Inspired by @karpathy's nanochat, we stripped down our prototype inference engine in pure Rust into a tiny hackable repo and open-sourced it (github.com/AntigmaLabs/nanochat-rs), as our tribute to the open-source community. And we are the first production-grade coding agent with first-class support for local models.

Alpha Preview​

Try it, currently on macOS and Linux. Expect breaking changes.

curl -fsSL https://ante.run/install.sh | bash

We'll share more in detail in the coming weeks. In the meantime — come try it and tell us what breaks.