Welcome: agents.erl! `new open source code release from the Arthur Collé Research Lab`

Hello, agents.erl

A minimal-but-mighty Erlang framework for OpenAI-powered agents, now open-sourced—and where we hope to take it next.

→ repo link ←

---

1. Why agents.erl? 🤔

Erlang/OTP was built for phone switches that never go down.
AI agents that juggle dozens of external calls (and occasionally crash) deserve the same reliability.

Agents.erl marries Erlang’s fault-tolerant supervision trees with OpenAI’s API so you can:

• Spin up thousands of concurrent agent processes (each ~kB in RAM).

• Let OTP (Open Telecom Platform) restart any process that blows up.

• Hide rate-limiting, streaming, and tool execution behind a clean API.

---

2. What’s in the repo right now?

Area Status Key files Dynamic API clients

✓ Auto-generated from the OpenAI spec (chat, completions, embeddings, etc.) apps/openai_clients/* Supervision tree

✓ One worker per API family + top-level supervisor agent_sup.erl Tool execution

✓ Register any Erlang function; agent can call it via function-calling messages agent_tools.erl Streaming support

✓ Handles delta chunks for chat completions openai_chat.erl Rate limiting Basic counter + sleep openai_rate_limiter.erl Example prompt loop

✓ Simple REPL cli/agent_cli.erl

🚧 What we don’t have (yet):

• No replication or clustering logic.

• No self-patching / hot-code helpers beyond standard OTP.

• No RL or neuro-evolution hooks.

• Repo still needs a proper Rebar3 umbrella layout and .beam / _build files cleaned out.

---

3. Lessons from the Hacker News launch

The Show HN thread (33 points, 21 comments) surfaced three concrete action items:

1. Remove compiled artefacts – shipping .beam files was… over-eager.

2. Rebar3 structure – a clean rebar3 new umbrella layout makes dependency management painless.

3. Elixir friendliness – keep the public API thin so Elixir apps can :rpc.call without caring about Erlang syntax.

All three are on the near-term roadmap.

---

4. Near-term roadmap (v0.2 → v0.3)

Version Goal ETA v0.2 Rebar3 umbrella, Hex.pm package, remove build junk, CI & dialyzer pass. ~2 weeks v0.3 Basic clustering: single ETS registry, agent_seed:clone/2 proof-of-concept, PromEx metrics. ~1 month

---

5. Long-term vision 🌌 (why this project excites us)

Below are aspirational ideas—not implemented—guiding the design:

1. Hot-patchable agents
   Agents propose Erlang source, compile it in a sandbox node, run tests, then hot-load it. (Think REPL-driven “immune system” for bugs.)

2. Mesh replication
   A tiny pg2-style mesh where agents migrate to the least-loaded node in <200 ms, streaming state deltas en route.

3. Self-evolving policies
   Plug a PPO loop + NEAT genome into each replica; good mutations propagate, bad ones die locally—continuous improvement without downtime.

4. Tool marketplace
   Agents publish JSON-schemas for their tools and barter calls based on latency & token cost.

5. Carbon-aware scheduling
   Heavier jobs slide to regions with greener energy by consulting CO₂signal.

These are research directions—code will land behind feature flags as prototypes harden.

---

6. How you can help 🙏

• Star or fork the repo → https://github.com/arthurcolle/agents.erl

• Open an issue for missing docs, build hiccups, or features you’d like.

• Share example tools or supervision patterns that make sense for agents.

• If OTP excites you but Erlang syntax doesn’t—help sketch an Elixir wrapper!

---

Closing thought

This release is intentionally small: a reliable, readable starting point.
If the idea of fault-tolerant AI agents running on the BEAM resonates with you, jump in—the road ahead is wide open, and we’d love collaborators on every rung of the ladder.

— Arthur (@arthurcolle)