Capabilities you can install

VA

Vercel AI SDK

Vercel

The library most JavaScript projects reach for when adding a model call, mostly because streaming, tool calls and provider swapping are handled before you write any of it.

L

LangChain

LangChain

The largest ecosystem and the largest surface area. It will have an integration for whatever you need; deciding how much of it to adopt is the actual work.

L

LlamaIndex

LlamaIndex

Narrower than LangChain and sharper for it: retrieval over your own documents is the whole product, and the parsing and chunking are where the quality comes from.

C

Composio

Composio

Pre-built authenticated connectors for the services agents usually need, which removes the least interesting week of any agent project.

C

CrewAI

CrewAI

Frames a task as a team of roles with a manager, which maps well onto how people already describe their work and badly onto anything where the roles are not really separable.

D

DSPy

Stanford

Replaces prompt writing with optimisation over examples, which is either the right answer or an over-engineering depending on whether you have examples. Worth reading the idea even if you never adopt it.

PA

Pydantic AI

Pydantic

Typed outputs as the organising idea, so a malformed answer fails at the boundary instead of three functions later. The choice for people who already live in type checkers.

S

smolagents

Hugging Face

Agents that act by writing code rather than by calling a tool schema, which is fewer moving parts and a much bigger security surface. Small enough to read in an afternoon.

SK

Semantic Kernel

Microsoft

The one to look at inside a .NET or Azure shop, because the integration with the rest of that stack is the point rather than the abstractions.

A

AutoGen

Microsoft

The research lineage of multi-agent conversation, with the most published work behind it and the most academic feel in its API.

H

Haystack

deepset

A pipeline framework that predates the current wave, and the maturity shows in the evaluation tooling rather than in the model integrations.

L

Letta

Letta

Built around persistent memory as the primary thing rather than as a bolt-on, which is the right starting point if your agent needs to remember last week.

M

Mem0

Mem0

A memory layer you add to an existing application rather than a framework you build on. The honest framing of what it saves is the plumbing, not the thinking.

RA

Relevance AI

Relevance AI

A no-code way to assemble an agent team, aimed at operations rather than engineering. The limit is the same as any visual builder, and it arrives later than you expect.

T

Toolhouse

Toolhouse

A hosted toolbox that an agent can call, so the tools are somebody else’s uptime problem. Worth the trade when the tools are peripheral and not when they are the product.

TM

The MCP server ecosystem

Community

The collection of published servers, from filesystem and git access to databases and issue trackers. Quality varies enormously, and the ones that wrap a database deserve a long look before you install them.

OA

OpenAI Agents SDK

OpenAI

Small, opinionated and built around handoffs between agents. It is a thin layer rather than a platform, which is a compliment and also the reason some teams outgrow it.