This presentation explores the current state and future of AI agents, particularly within the Elixir programming ecosystem. The speaker, Chris, argues that Elixir, with its foundation in OTP and the BEAM VM, is uniquely and "accidentally" perfectly suited for developing and leveraging these powerful AI tools.
The talk features a live demonstration where an AI agent, powered by Cloud Code and guided by a custom agents.md file, successfully builds a functional Elixir Phoenix Slack clone application, showcasing real-time features and robust OTP architecture. Chris emphasizes practical agent development strategies, including "gap-filling" LLM knowledge with targeted instructions, simulating conversational flows, and even manipulating agent memory for error correction. He also addresses common debates surrounding AI's impact on productivity and reframes the narrative to highlight Elixir's opportunity to attract new developers by seamlessly integrating with powerful AI coding assistants, thus lowering the barrier to entry. The presentation concludes with a call for the Elixir community to embrace this agentic age and lead in building both with and for AI agents.
Summarizing "Elixir and the Agentic Age"
0:15 AI Agents: The Future of Programming: The speaker introduces AI agents as the current "rage" in tech and programming, fueled by their ability to perform complex tasks in a loop.
0:53 Live Demo: AI Builds an Elixir Slack Clone: A demonstration begins with an AI agent (Claude via cloud code) generating a Phoenix LiveView Slack clone application.
1:19agents.md for LLMs: Phoenix 1.8 projects include an agents.md file, designed to "smooth over" LLM knowledge gaps (e.g., Elixir-specific syntax or patterns) to improve agent coding experience, rather than serving as human documentation.
6:23 Exponential Growth in Agent Task Capacity: Research indicates agent task length capacity is doubling every seven months, suggesting a rapid increase in their ability to stay on track and complete longer, more complex tasks.
8:10 Industry Predictions for AI Agents: Leaders from Amazon and OpenAI predict billions of AI agents integrating into every company and materially changing workforce output by 2025, moving beyond simple chat assistance to collaborative, independent work.
12:11 Elixir's Accidental Suitability: Elixir (and the BEAM VM) is presented as the "perfect language" for the agentic age, similar to how it was accidentally well-suited for multicore processing, the internet, and concurrent systems before they became widespread.
13:51 Ignore the AI Hype Debate: The speaker advises against getting caught up in the "doom vs. utopia" debate surrounding AI, suggesting developers instead focus on leveraging tools to "ship faster than ever thought possible."
15:11 Elixir's History of Bleeding-Edge Innovation: Elixir's consistent focus on building the "future we want to see" (e.g., Phoenix Channels, GenStage, Presence, LiveView, NX) demonstrates it's not "chasing" hype but rather leading in technological advancements.
19:09 Elixir's Dev-First Tooling Aids LLMs: Elixir's cohesive, developer-first tooling (Mix, collocated documentation, annotations like at_collaller and data-phx-line) inherently provides LLMs with better context and understanding of the codebase.
21:38 Reframing the "Elixir is Irrelevant" Argument: The speaker counters the fatalistic view that LLMs, being trained on JavaScript, will make Elixir irrelevant. Instead, he argues that if "language doesn't matter," LLMs can lower the barrier to entry for new Elixir developers, allowing them to instantly generate functional Phoenix apps and grow the community.
35:11 Agents are Trivial: Simulating a Chat: The core of an AI agent is simply a recursive loop simulating a chat with an LLM. "Function calls" or "tool calls" are just the LLM generating text in a format the program understands to dispatch to real-world functions.
39:07 Hidden Messages and Memory Rewriting: Advanced agent techniques involve inserting "hidden messages" (e.g., promises from the agent) into the chat history to influence behavior and rewriting the LLM's memory to correct errors, making it appear as if the agent produced the correct code initially.
47:46 Keep Agent Architecture Simple: Advice to focus on a single main thread for agents, avoiding complex "sub-agent" trees, due to increased failure modes and the rapid improvement of frontier models.
50:11 Agents Writing Code That Writes Code: Exploration of agents writing small Lua scripts to perform partial file edits, effectively writing custom parsers in real-time. This method is 80% successful and represents a future direction for agents.
56:03 Demo Success & Key Takeaway: The live demo successfully produces a working Slack clone with presence, pub/sub, and GenServer persistence. The speaker concludes that Elixir is perfectly positioned to dominate the agent space, enabling developers to build powerful, production-ready applications in minutes.
59:44 Q&A: Local vs. Frontier Models: In response to a question, the speaker advises against fine-tuning and using local models for most use cases currently, citing their lower capability and the rapid advancements and cost-effectiveness of frontier cloud models (e.g., Gemini Flash).