Upsonic

Upsonic splits agents into two shapes. Autonomous agents handle open-ended, multi-step work and can touch the filesystem and shell inside a locked workspace. Traditional agents fit structured tasks with a defined input and output contract. You pick based on the problem, and both run through the same pipeline.

You write a function, mark it with a decorator, and hand it to an agent as a tool. Beyond your own code, Upsonic speaks the Model Context Protocol, so agents reach thousands of external services without custom glue. Ready-made connectors cover search, crawling, sandboxes, and finance data.

Agents keep conversational memory and pull from a RAG knowledge base for retrieval. Store that memory in SQLite, Postgres, Redis, Mongo, or Mem0, and build retrieval pipelines on Qdrant, Pinecone, Chroma, pgvector, or Weaviate. The point is continuity: an agent that recalls earlier context beats one that starts fresh each run.

A Team groups several agents and runs them in sequential, coordinate, or route mode. Each agent gets its own name, role, and goal. This suits work that genuinely needs different specialties, like one agent researching and another writing, rather than a single agent pretending to do everything.

The OCR layer turns PDFs and images into text through a layered pipeline, with engines including EasyOCR, Tesseract, PaddleOCR, and DeepSeek. It matters for teams whose data lives in scans and invoices rather than clean text files.

Guardrails screen content against preset privacy, financial, and security policies. Human-in-the-loop checkpoints pause a run for approval or input. Tracing hooks into Langfuse and PromptLayer, so you can see the path an agent took, not just its final answer.