pip install deepagents pulls langchain-anthropic and langchain-google-genai (google-auth, cryptography, ...) even if you only use Ollama; slow container builds
Expect pip install deepagents to be a big install: deepagents 0.7.38 hard-requires langchain-anthropic>=1.7.10 and langchain-google-genai>=4.4.3 (plus langchain, langchain-core>=1.8, langsmith, wcmatch). Put it in its own cached layer so you don't re-download it on every container build. Through google-genai you also get google-auth, pyasn1-modules, cryptography and friends. On a busy machine our podman build with pip install --no-cache-dir langchain langchain-ollama langgraph deepagents ... took about 5 minutes and outlived a 3-minute build timeout, even though we only talk to a local Ollama model. Fixes: split RUN pip install deepagents langchain-ollama into its own layer above COPY . /app so code changes don't rebuild it, use a BuildKit/podman cache mount instead of --no-cache-dir, and pin versions (deepagents==0.7.38 resolved langchain 1.5.0, langgraph 1.2.22, langchain-ollama 1.1.0). Why Anthropic is a hard dependency: create_deep_agent(model=None) falls back to ChatAnthropic(model_name="claude-sonnet-4-6"). That default has been deprecated since 0.5.3 and goes away in 1.0, so pass a model explicitly, e.g. ChatOllama(model="qwen2.5:3b", num_ctx=8192). Extras such as deepagents[aws] and [video] are optional; the Google and Anthropic integrations are not.