A Python agent framework that thinks observability is not optional

What it does

Datapizza AI is a Python framework for assembling LLM agents, retrieval-augmented generation pipelines, and document-ingestion workflows. It wraps multiple providers—OpenAI, Anthropic, Google Gemini, Mistral, and Azure OpenAI—behind a uniform client interface and lets you chain them into directed-acyclic-graph pipelines or delegate tasks across a hierarchy of specialized agents. The pitch is less magic abstraction and more explicit control over prompts, tools, and execution flow.

The interesting bit

Most frameworks treat tracing as an afterthought; Datapizza AI emits OpenTelemetry spans by default and will even print a tidy CLI summary of token counts, latencies, and cached hits for every run. That focus on observability is the unusual angle—it is trying to solve the “why did my agent do that?” problem before you ship to production.

Key highlights

• Modular provider support: Clients for OpenAI, Gemini, Anthropic, Mistral, and Azure are packaged as separate, optional modules.
• DAG pipelines for RAG: The DagPipeline class wires query rewriters, embedders, retrievers, and generators with explicit connect() calls.
• Agent delegation: Agents can hand off to other agents via can_call(), enabling multi-agent setups like a trip planner that consults a weather expert and a search expert.
• Document ingestion stack: Parsers for PDF and DOCX (via Docling or Azure AI), chunking, and embedding routes into Qdrant vector stores.
• Observability by default: OpenTelemetry tracing, optional client I/O logging, and custom spans are built in, not plugin-based.

Verdict

Worth a look if you want explicit control over Python agent execution with tracing included out of the box. If you need support for vector stores beyond Qdrant or a fully managed no-code platform, keep looking.