Python Agent Frameworks

What's in this pack

This pack collects the five Python-first agent frameworks that ship enough production code to use today, ranked by GitHub stars and active commit cadence. Each one represents a different design philosophy — pick by the shape of your problem, not the brand name.

The list intentionally excludes LangGraph (covered as a default in most LangChain stacks) and the JS-only frameworks. It covers the design space Python users actually pick from when LangGraph's StateGraph feels too heavyweight.

Why "stay in Python" matters

Most teams have a Python data stack already — pandas, FastAPI, Postgres drivers, the ML libs. Switching to a TypeScript or Go agent runtime means rewriting the data plumbing for one feature. These five frameworks sidestep that entirely. They run inside your existing FastAPI service, share your venv, and call the same DB driver your backend already uses.

The trade-off shape:

• Phidata is the most "Python-y" — agents are classes with tool methods, very SQLAlchemy-flavored. Best when your agent IS a data app (dashboards, internal tools).
• AGiXT is a full platform with a UI, provider abstraction, chains, extensions. Heaviest install, most features out of the box.
• AutoGPT is the autonomous-loop archetype — goal in, plan + execute + reflect, repeat. Heavier on token cost; lighter on dev work.
• OpenAI Swarm is the smallest — agents hand off to each other via a transfer_to_X() function. Experimental but the cleanest mental model for routing.
• CrewAI lets you define Agent, Task, Crew objects — explicit roles, sequential or hierarchical execution. Great for "marketing → review → publish" pipelines.

Install in one command

# Install the pack — drops requirements + example agents into your project
tokrepo install pack/python-agent-frameworks

# Or pick one
tokrepo install crewai
tokrepo install openai-swarm
tokrepo install autogpt

The TokRepo CLI fetches each framework's getting-started template, places it under agents/<framework>/, and adds dependencies to pyproject.toml or requirements.txt. Run pytest agents/ to validate the examples before wiring to your real prompts.

Common pitfalls

• Don't use AutoGPT for closed-form tasks. Its strength is open-ended goals. For "summarize this PDF" you'll burn tokens reflecting on a problem that has one obvious step.
• Swarm is experimental, not production. OpenAI ships it as a "design pattern" repo. Use the routing pattern; build the production loop yourself or copy it into a maintained fork.
• CrewAI sequential mode hides parallelism. If two tasks have no dependency, declare them parallel — sequential mode runs everything end-to-end even when not needed.
• Phidata's storage layer assumes Postgres. SQLite works for local dev but the storage interface is built around JSONB. Plan a real Postgres for any deployed run.
• AGiXT updates fast. The schema between minor versions has broken before. Pin a specific tag in production and read the changelog before upgrading.

When this pack alone isn't enough

If your problem is one Python-only service, this pack is enough. If you have:

• a Java/Spring backend → see Agent Frameworks Across Languages for Spring AI and LangChain4j
• a TypeScript edge function → same multilang pack covers Mastra
• a non-LLM evaluator gating step → pair with LLM Eval & Guardrails

You can also combine: a CrewAI orchestrator written in Python that calls a Spring AI agent in Java for a specific tool, evaluated by Promptfoo before shipping. The frameworks here are agent runtimes — they don't dictate the rest of your stack.