Rankings
aaaa.fyi · 2026-07-23 · 7 min read
A framework earns its place by handling what you'd otherwise rebuild: the agent loop, typed tool calling, retries, memory, and tracing. Here's how the field looks — grounded in the GitHub activity data we track for every framework in the directory.
Graph orchestrators. Model your agent as a graph of steps with explicit state. Most control, most ceremony. The right shape for production systems where you must audit every transition.
Multi-agent frameworks. Several specialized agents cooperating — a researcher, a coder, a critic. Powerful for fan-out work; watch the token bill and the coordination complexity.
Lightweight SDKs. A thin, typed layer over the model API with tool-calling and handoffs. The fastest path from idea to working agent, and where much of the ecosystem's momentum moved.
No/low-code builders. Visual pipelines for teams that need business users in the loop.
MCP standardized tools and skills standardized expertise — the parts frameworks used to differentiate on. The loop itself is commoditizing. Pick the *thinnest* framework that meets your constraint, and put your effort into tool design, which is where agents actually fail.
Compare real activity (stars, forks, last push): agent frameworks, ranked.