1.0.0 (2026-11-16)
⚠ BREAKING CHANGES
- core: passing an option an architecture does not take raises TypeError.
- llm: requires langchain-core, langchain and langgraph 1.x. The default vector store is the in-memory one, and LLM_MODEL no longer defaults to Llama 3.3 70B for every provider.
Features
- add DeepSeek provider + Mixture of Agents (MoA) architecture (#18) (377f57e)
- notebooks: notebook 36, Mixture of Agents, built and executed (14419ce)
- tools: SEARCH_BACKEND=mcp searches through a keyless MCP server (db5721b)
- tools: SEARCH_MCP_API_KEY for an MCP search server's key (53b887a)
Bug Fixes
- agents: MemGPT and SWEAgent require the arguments of their actions (5458d90)
- agents: SWEAgent and MemGPT tell the model what is missing instead of failing quietly (347cf6c)
- config: an empty value in .env means not set (f20e476)
- core: an unknown architecture option is an error, not ignored (85b37b9)
- import the LLM factory lazily, which the settings' provider lookup needs (82979a7)
- llm: each provider gets its own default model, and Google, Ollama and embeddings work (401a423)
- llm: give Ollama a 16k context window by default (9c807e0)
- llm: Ollama supports structured output (21ec15a)
- llm: replace the default models providers have retired (7ec1bf9)
- moa: each layer reads only the layer before it, as in the paper (d970124)
- notebooks: stop hardcoding Nebius models that have been retired (a7d7d14)
Documentation
- Mixture of Agents in the listings, keyless search and a free local path in the README (f727f65)
- sync the site's notebook copies with notebooks/ (18586d9)
- the November 2026 models, keyless search, OpenRouter and OpenAI-compatible servers (126b127)
- the test numbers, and what the real-model sweep showed (7649699)
Tests
- integration: five tests that could not fail or could not pass (d1692ce)
- integration: INTEGRATION_LLM_KWARGS passes extra model arguments to the sweep (e936899)
- integration: the CellularAutomata grid is one string literal (70b8230)
- providers, embeddings fallback, the in-memory store, search backends and MCP tools (9b8c13f)
- the integration sweep retrieves with the real embeddings model (56b66b8)