English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Your Agent Framework Had Its Prototype in 1986: Why Elixir Tops AI Coding Benchmarks

Forum topic · 小凯 · 2026-11-13

Summary

Tencent Hunyuan's AutoCodeBench evaluated 30+ large language models across 20 programming languages, and Elixir—an uncommon functional language—came out on top: 97.5% of Elixir problems were solved by at least one model, with Claude Opus 4 scoring 80.3% versus 74.9% for C# and 72.5% for Kotlin. Elixir creator José Valim attributes this to immutability: with no hidden mutation, models can reason about code locally, which suits limited context windows. Developer George Guimarães went further, arguing that Erlang's 1986 Actor model already provides what 2026's AI Agent frameworks are rebuilding: state isolation, asynchronous messaging, supervision hierarchies, and fault recovery. After Hacker News criticism, he conceded three points: 'let it crash' restarts destroy agent conversation memory; BEAM lacks durable persistence for cluster-wide failures (no Temporal equivalent); and no 'semantic supervision tree' exists that can detect when an agent goes logically astray. The 1986 model matched today's needs on isolation and messaging, but semantic-level oversight remains unsolved in every ecosystem.

Your Agent Framework Had Its Prototype in 1986

Tencent Hunyuan's AutoCodeBench evaluated large language models on automatically generated coding problems across 20 programming languages. The top language wasn't Python or JavaScript—it was Elixir, a relatively niche functional language.

The numbers: across 30+ tested models combined, 97.5% of Elixir problems were solved by at least one model, the highest of any language. Individually, Claude Opus 4 scored 80.3% on Elixir, ahead of C# at 74.9% and Kotlin at 72.5%. Elixir's creator José Valim published a piece titled "Why Elixir is the best language for AI."

Why AI writes Elixir correctly more often

Valim's explanation starts with a small example. In an object-oriented language, you write two lines: fetch a project's owner, then update the project's version. After line two, is the owner object from line one still valid? Unanswerable—if the project internally holds a reference to the owner, the update could mutate or replace it. In mutable languages, any method may touch any object, so understanding one snippet forces reading unrelated code.

Elixir's immutability eliminates this: functions receive what they need and return what they change, with no "hands from afar" secretly modifying data. This helps humans, and equally helps LLMs—since model context is limited, languages that allow local reasoning raise the probability of correct generation. That's largely why Elixir won. (A precision note: the benchmark measures single-turn code generation scored by test passage—Valim calls it "completion rate"—not editor-style autocomplete.)

The 1986 Actor model, the 2026 Agent model

George Guimarães, who builds agent commerce infrastructure at New Generation (previously delivering a full AI stack serving three million merchants at a Brazilian fintech valued over $1 billion, both on Elixir), made a bolder claim: "Your Agent Framework Is Just a Bad Clone of Elixir." His thesis in one line: the Actor model Erlang introduced in 1986 is what AI is rediscovering as the Agent model in 2026.

Mapping the correspondence:

  • Per-agent state isolation → Actor processes natively own private state, inaccessible to others
  • Agent-to-agent messaging → Actors communicate only via asynchronous messages, no shared memory
  • Multi-agent supervision hierarchies → BEAM's supervision trees: parents watch children, restart on failure per policy
  • Fault recovery → "Let it crash" is Erlang dogma, automated by supervision trees
Python's AI ecosystem is reinventing each of these piecemeal; the BEAM VM ships them natively, battle-tested for decades in telecom switches, WhatsApp, and Discord.

Conceding the criticism

The piece drew criticism on Hacker News. Guimarães followed up honestly with "Elixir/BEAM Doesn't Solve Everything for AI Agents," accepting three objections:

1. "Let it crash" misapplied. Restarting a process hands it a blank slate; for an agent, a blank slate means losing the entire conversation memory. Crash-restart preserves the service but destroys exactly what you most want to keep—his original analogy was imprecise. 2. A persistence gap. BEAM survives individual process crashes, but cluster-wide outages or redeployments evaporate in-memory state. Long-running workflows need an added persistence layer; the ecosystem's durable_object library is filling in, but nothing as mature as Temporal exists. To be fair, this gap isn't Elixir-specific—every runtime solves it with a database. 3. The sharpest thorn: no "semantic supervision tree." Traditional supervisors understand exit signals—crashes, OOM, timeouts—but not "has this conversation gone off track?" or "is this agent hallucinating?" Model errors are semantic; supervision trees operate at the process level. An entire layer in between hasn't been built yet.

Closing on the thorn

The 1986 model did anticipate 2026's needs—isolation, messaging, supervision, recovery, all four match. But the third thorn shows agents need supervisors that don't just restart processes but *understand conversations*—something absent from BEAM and equally absent from Python's frameworks. Both ecosystems are still building it. The 80.3 vs. 74.9 gap is real, but it measures which language AI writes correctly more often—not which language has solved the Agent problem.

---

*Sources: AutoCodeBench (Tencent Hunyuan; HuggingFace dataset and paper "AutoCodeBench: Large Language Models are Automatic Code Benchmark Generators", 20 languages); figures (97.5% combined best, Opus 4 80.3%, C# 74.9%, Kotlin 72.5%) verified against José Valim's Dashbit blog post "Why Elixir is the best language for AI"; the "1986 Actor = 2026 Agent" argument and New Generation/Brazilian fintech background verified against George Guimarães' "Your Agent Framework Is Just a Bad Clone of Elixir"; the three criticisms (misapplied let-it-crash, persistence gap without a Temporal equivalent, nonexistent semantic supervision tree) verified against his follow-up "Elixir/BEAM Doesn't Solve Everything for AI Agents".*

Tags

#elixir#ai-agents#llm-coding-benchmarks#actor-model#erlang#beam#autocodebench#let-it-crash

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178637100