How to pick an AI agent framework in 2026
2026-11-13
How to pick an AI agent framework in 2026
Two hundred starred repos describe themselves as "AI agent frameworks". Almost
none of them will still matter in a year. Here is a filter that takes an
afternoon, not a quarter.
1. Start from the runtime you already have
A framework is a bet on someone else's abstraction. If your team already runs
TypeScript, a Python-first framework means a second deploy pipeline for every
agent you ship. The collector's directory lets you filter by language for exactly
this reason — the popular choice is usually the one that matches your stack.
2. Count the things you actually need
Most agent projects need some mix of: tool calling, retrieval (RAG), multi-agent
orchestration, human-in-the-loop, and observability. Write your list before you
compare frameworks; feature grids written afterwards always favor the one you
just installed.
3. Check who pays when it breaks
- License: permissive (MIT / Apache-2.0) vs copyleft (AGPL). This is a
legal question, not a technical one.
- Release cadence: a framework with no commits in 90 days is a prototype.
- Lock-in: can you swap the model provider without rewriting the agent? If
the framework hardcodes one vendor's request shape, every future price cut
from a competitor costs you a rewrite.
4. Model choice is a separate decision
Framework and model are independent. A good framework lets you point the agent
at any OpenAI-compatible endpoint; a bad one quietly assumes one vendor. This is
the same reason our gateway forwards requests without rewriting the body — the
model layer should be replaceable on a Tuesday afternoon.
5. Pilot on your worst case
The honest test is not the demo from the framework's landing page. It is your
messiest real task: the long context, the weird tool, the flaky upstream. Run
the same task through two frameworks and read the traces.
What we look for in the directory
The agent directory indexes 130+ repos with
framework, language, intent tags and star counts, so you can run this filter in
public. See also choosing a model for your workload.