Upsonic

Upsonic's bet is that the hard part of agents is not orchestration syntax but containment. Every AutonomousAgent runs inside a workspace, and every file and shell operation is restricted to it — path traversal and dangerous commands are blocked at the framework level, not by a prompt. That single design decision is why the framework is worth reading the docs for, and it is the thing most competing libraries leave to you. The second primitive, Traditional Agent, is the escape hatch for problems that were never open-ended: you define a @tool, a response schema, and a Task with an explicit input/output contract, and you get deterministic plumbing where you want it. The supporting cast is unusually complete for a library at this stage. Memory is pluggable across SQLite, Postgres, Redis, Mongo and Mem0. Knowledge Base covers the whole RAG path — loaders, splitters, 7+ embedding providers, and vector stores including Qdrant, Pinecone, Chroma, pgvector, Weaviate and Milvus. The v0.75.0 release hardened this surface with a KBState enum (UNINITIALIZED → CONNECTED → INDEXED → CLOSED), a KnowledgeRow table with CRUD across all seven storage providers, content-hash deduplication that skips unchanged documents and deletes stale chunks before re-indexing, and an isolate_search flag so multiple KnowledgeBases can safely share one collection. That is the kind of work that only matters in production, and it is a good signal about who this is built for. Teams coordinate through Sequential, Coordinate or Route modes. Human-in-the-loop gives you confirmations, user input, and durable execution pauses so an agent that touches a shell can ask first. Skills package reusable capabilities loaded from local paths, URLs or GitHub. MCP support connects external tool servers, and the ready-to-use integrations cover Tavily, Firecrawl, E2B and YFinance. Observability is handled through Langfuse and PromptLayer rather than reinvented. The Applied Scientist agent is the clearest product bet. It automates the boring, error-prone first week of a paper: you point research_source at a PDF, a Markdown file, a .ipynb, a blog post, an arXiv page, a git repo, a Kaggle notebook, a Hugging Face model card, or just type the idea, and the agent detects the source kind at Phase 0, materializes it into the experiment folder, and runs the experiment against current models. Since v0.76.2 you no longer pass inputs=[...] by hand — anything pointing at an existing local path is auto-copied — and since v0.76.3 you can drop current_data entirely and let the agent read the current notebook to infer the data source. Where it doesn't fit: this is a library, not an app. There is no chat UI, no dashboard, no control plane. You install via pip, own deployment, scaling and provider key configuration, and wire your own front end. The release cadence is fast — v0.76.1, v0.76.2 and v0.76.3 all landed on Apr 22, 2026 — and APIs are still moving (the Applied Scientist's research_paper input was renamed