I have 5+ years of commercial experience, including 4 years as a Software Engineer at Software Mansion in Kraków, Poland. Currently contracting for a large U.S.-based company, working primarily with Elixir. On a day-to-day basis, we work with an architecture handling tens of thousands of active users at a time, leveraging the strengths of BEAM, Phoenix, Ecto, and Oban. We’re actively incorporating AI into our workflows. Read more about my experience on my blog: dimamik.com/about.
Most AI agents are trapped in a “chat-and-call” loop, restricted by rigid function-calling interfaces. Legion breaks this cycle. It introduces a paradigm where agents don’t just request actions - they execute them. By allowing LLMs to run vetted Elixir code in isolated sandboxes, Legion enables agents to spawn other agents, exchange messages, and react to events in real-time. There is no new DSL or steep learning curve, Legion turns your pure modules and functions into an autonomous toolbox. You write the limbs - Legion provides the nervous system to orchestrate them.
https://github.com/dimamik/legion
Key Takeaways:
Beyond RPC: Why executing full code snippets creates faster, more capable and efficient agents than single-action function calls.
Authorization: Why Vault (https://github.com/dimamik/vault) is the most robust way to authorize tool calls at scale.
The Elixir Advantage: Why the BEAM’s isolation and supervision are uniquely suited for autonomous AI systems.
Target Audience: