Axocoatl: Rust Runtime for Persistent AI Agents

Axocoatl Release on GitHub

Axocoatl appeared on GitHub Trending as a Rust-based agentic runtime for persistent, supervised agents. The project, hosted at

, targets self-hosted and local-first deployments with zero telemetry and an Apache 2.0 license. It implements multi-agent coordination through a stigmergic event lattice instead of a central controller. Agents activate on pheromone-style signals once dependencies finish, and the system supports checkpoint restarts after interruptions such as a killed server process.

Core Technical Features

Axocoatl builds on the ractor actor model for low memory use and fast cold starts. Each agent maintains four tiers of persistent memory and can resume from the last checkpoint rather than restarting tasks from scratch. The runtime includes HTN symbolic planning, auction-based agent selection, and per-agent token budgets. It also ships an MCP client and server plus A2A protocol support.

Coordination happens without an orchestrator. Agents place signals on the lattice, and dependent agents wake when conditions match. This stigmergic approach differs from explicit task graphs used in Python frameworks. The configuration file defines agents, workflows, and scheduled runs in a compact YAML format that fits the entire system definition on one screen for small demos.

Provider support covers Ollama, OpenAI, Anthropic, and similar endpoints through a single abstraction layer. The daemon exposes an API for chat sessions and monitoring while remaining runnable entirely offline.

Setup Process

Installation starts with a shell script that places the CLI binary without requiring a Rust toolchain on the host machine. After running the onboard wizard, users receive a scaffolded project and chosen provider configuration. The doctor command checks environment variables, model availability, and lattice connectivity before launch.

Starting the system uses two commands: one for the daemon and API server, another for interactive chat with a named agent. Developers who prefer Cargo can install the CLI crate directly. The repository includes both a minimal two-agent example and a larger twelve-agent demonstration with scheduled runs and MCP server definitions.

Trade-offs and Limitations

Rust and the actor model deliver stronger isolation and lower overhead than Python-based alternatives, yet the ecosystem lacks the breadth of pre-built tools available in CrewAI or AutoAgents. The stigmergic lattice requires explicit dependency modeling in YAML, which can lengthen initial setup compared to imperative task lists.

Checkpointing works across restarts only when the underlying storage layer remains intact. Provider-agnostic calls still expose model-specific token limits and latency differences that the runtime does not abstract away. Early adoption means fewer community extensions and less battle-tested behavior under sustained high agent counts.

The project prioritizes local execution and auditability over managed cloud features. Teams already running Rust services may find integration straightforward, while those invested in Python agent frameworks face a language switch for the runtime layer.

FAQs

Does Axocoatl require a Rust compiler for everyday use? No. The install script delivers prebuilt binaries, though Cargo installation remains available for those who want to build from source.

How does the stigmergic lattice differ from standard task orchestration? Agents react to signals left by completed peers rather than following a central scheduler, removing the need for an explicit orchestrator process.

Can Axocoatl agents survive a full server reboot? Yes, provided the persistent storage layer holds the checkpoint data; agents reload state and resume at the last recorded step.

---

📖 Related articles

• Agentic Coding: Una Trappola per lo Sviluppo Software?
• File agents.md: utili per gli agenti di coding?
• Lean-ctx: Ottimizzatore Ibrido Riduce Consumo Token LLM del 89-99%