Alloy 0.8.0 - Minimal, OTP-native agent loop for Elixir

Alloy 0.8.0 The model-agnostic agent engine for Elixir

Alloy is a minimal agent loop for building AI agents on the BEAM. Think of it as the completion-tool-call loop and nothing else, no opinions on sessions, persistence, or UI.

What’s new in 0.8.0

Cost guard pass max_budget_cents: 50 to Alloy.run/2 and the loop halts with :budget_exceeded status before your agent overspends. The check runs at the start of each turn alongside max_turns.

{:ok, result} = Alloy.run("Research this codebase",
  provider: {Alloy.Provider.Anthropic, api_key: key, model: "claude-sonnet-4-6"},
  tools: [Alloy.Tool.Core.Read, Alloy.Tool.Core.Bash],
  max_budget_cents: 50
)
# result.status => :budget_exceeded | :completed

Summary-based context compaction when conversations approach the token limit, Alloy generates a structured handoff summary of older context instead of truncating. The summary preserves progress, key decisions, constraints, and next steps. Falls back to deterministic truncation if summary generation fails.

Compaction is now configurable:

Alloy.run(input,
  provider: provider,
  compaction: [
    reserve_tokens: 12_000,       # Room for the next response
    keep_recent_tokens: 8_000,    # Recent messages kept verbatim
    fallback: :truncate           # When summary fails
  ]
)

One-shot streaming Alloy.stream/3 gives you token-by-token output without managing a GenServer:

{:ok, result} = Alloy.stream("Explain OTP", fn chunk ->
  IO.write(chunk)
end, provider: {Alloy.Provider.OpenAI, api_key: key, model: "gpt-5.4"})

Why model-agnostic matters

Alloy works with Anthropic, OpenAI (Responses API), xAI/Grok, Gemini, and any OpenAI-compatible endpoint (Ollama, OpenRouter, DeepSeek, Mistral, Groq). Swap providers by changing one line — your tools, prompts, and agent logic stay the same.

Today’s best model might not be tomorrow’s. Building on Alloy means you can switch without rewriting your agent.

The philosophy

Inspired by Pi Agent’s minimalism: 4 built-in tools (read, write, edit, bash), a good prompt, and a loop. That’s the framework. ~5,000 lines, small enough to read and understand in an afternoon.

OTP gives us the rest for free: supervision trees for fault tolerance, Task.Supervisor for parallel tool execution, PubSub for event streaming, GenServer for stateful agents.

Links

• Hex: alloy | Hex
• GitHub: GitHub - alloy-ex/alloy: Model-agnostic agent harness for Elixir · GitHub
• Docs: alloy v0.12.2 — Documentation
• Changelog: alloy/CHANGELOG.md at main · alloy-ex/alloy · GitHub

Feedback, issues, and PRs welcome. Happy to answer questions about the architecture.