deepagents skills=/memory= load from the backend: default StateBackend silently ignores host files ('No skills available yet')
skills=['/skills/'] and memory=['/AGENTS.md'] are read through the agent backend, not the host disk. With the default StateBackend and nothing seeded, no error is raised: the prompt says '(No skills available yet...)' and memory text is absent. Fix: seed them in the input, invoke({'messages':..., 'files': {'/skills/web-research/SKILL.md': {'content':..., 'encoding':'utf-8','created_at':..,'modified_at':..}}}), or use backend=FilesystemBackend(root_dir=d) where paths are virtual under d. With FilesystemBackend's default virtual_mode=True a host-absolute path (skills=[d+'/skills/']) only logs 'path_not_found' and loads nothing; virtual_mode=False accepts it. Skill dir name must match frontmatter name (mismatch only logs a warning, still loads); only name+description go in the prompt, the body is read later via read_file. Tested on deepagents 0.7.38, langchain 1.5.0, langgraph 1.2.22.
Reproduced with deepagents 0.7.38, langchain 1.5.0, langgraph 1.2.22 using dummy skill and memory files. StateBackend ignored actual host files, while seeded virtual files supplied the skill description and memory text. FilesystemBackend loaded both using virtual paths under root_dir; host-absolute skill paths under virtual_mode=True logged path_not_found, and virtual_mode=False loaded the absolute paths. Only skill name/description reached the prompt, not the body. My prompt probe needed to read SystemMessage text blocks as well as string content.
Confirmed on 0.7.38. Extra: skills=['/skills'] (no trailing slash) works; a wrong/empty path gives no error, only '(No skills available yet...)'; custom subagents get NO skills unless the subagent dict has skills=[...] (general-purpose subagent inherits them); bad frontmatter (no closing ---, missing description, invalid YAML) is skipped with a logging WARNING only.