Open-source AI agent frameworks and orchestration

Agent frameworks give you the loop — planning, tool calls, memory, retries — so you build the task instead of the plumbing. This is the deepest category in the catalog, which makes choosing harder than finding.

The single all-purpose assistant is being replaced by small teams of narrow agents that hand work to each other, and the frameworks below are the layer that makes that handover survivable: shared memory, a tool contract, and a defined behaviour when one step returns nonsense.

What separates them in practice is rarely the planner. It is whether state persists across a crash, whether you can inspect why an agent chose a tool, and whether extending it means writing a plugin or maintaining a fork — questions the READMEs answer last, which is why each project here is described by what it does rather than by how it markets itself.

Projects in this topic

87 analysed projects match this topic.

Developer Tools
Featured

Developer-first agent tools layer for claude

Skip weeks of custom work by starting from a working agent tools foundation.

agent toolsclaudedeveloper tools
Beginner · AI application
TOOL-76566View details
Developer Tools
Featured

Developer-first algorithmic trading layer for llm

Skip weeks of custom work by starting from a working algorithmic trading foundation.

algorithmic tradingllmmulti-agent system
Beginner · AI application
TOOL-86574View details
Browse the full catalog

Questions people ask

Do I need a framework to build an AI agent?
No. A single-tool agent is a loop, a prompt and an API call, and writing it yourself takes an afternoon. Frameworks pay for themselves at the point where you need persistent memory, several tools with conflicting failure modes, or more than one agent sharing state.
What is the difference between an agent framework and an orchestration layer?
A framework gives one agent its decision loop and tool interface. An orchestration layer coordinates several of them: routing, hand-offs, shared state and the policy for what happens when one agent stalls. Several projects in this list do both, which is worth checking before you combine two.
How much work is it to put an open-source agent into production?
The demo is a day; the production version is the observability, the cost ceiling and the fallback when the model returns malformed output. Every entry here carries a difficulty rating that reflects the second number, not the first.

Related collections

All collections

Other topics

Updated: 2026-08-17