| Type | Open source | Open source | Open source |
|---|
| CategoriesDiffers | | | |
|---|
| In one line | Low-level orchestration for controllable, stateful agents. | Role-based multi-agent teams with a simple mental model. | A lightweight, production-ready multi-agent SDK with tracing. |
|---|
| Stage & funding | — | — | — |
|---|
| GitHub starsDiffers | ≈44k | ≈60k | ≈30k |
|---|
| License | MIT | MIT | MIT |
|---|
| Language | Python | Python | Python |
|---|
| Last pushDiffers | Oct 28, 2026 | Oct 28, 2026 | Oct 27, 2026 |
|---|
| Works with | — | — | — |
|---|
| Built onDiffers | — | — | |
|---|
| Built on itDiffers | | — | |
|---|
| Similar toDiffers | — | | — |
|---|
| People linkedDiffers | 3 | 0 | 0 |
|---|
| EventsDiffers | | 0 | 0 |
|---|
| Pros | - +User reportsUsers running it in production cite checkpointing, pausing mid-graph for human input and time-travel over state as the main reasons to use it.[1][2][3]
- +User reportsUsers find the explicit graph/state-machine model a useful way to control agent flow, even when they dislike the wider LangChain abstractions.[1][2]
- +User reportsThe Postgres checkpointer lets stateless backends resume sessions without session affinity.[1]
- +MetadataMIT licensed, actively released (1.2.18 on 2026-10-26), about 44k stars.[1]
| - +User reportsUsers find it fast for prototyping, with role-based agents that are intuitive to set up.[1]
- +User reportsSeparate agents, tasks and tools, plus YAML config for roles and models, keep behaviour changes out of code.[1]
- +MetadataMIT licensed and released very often (1.15.33 on Oct 29, 2026).[1]
| - +User reportsVery easy start for function calling and sub-agents-as-tools; with the Responses API you don't have to store input items yourself.[1]
- +User reportsOne long-time user says they are largely happy with its abstractions: handoffs, tools, guardrails and structured output.[1][2]
- +DocsProvider-agnostic despite the name (100+ LLMs), with built-in tracing, sessions, sandbox agents and realtime/voice agents.[1]
- +User reportsCommenters see it as simpler than heavier frameworks like Google ADK or LangChain, and its backlog is small (about 12 open issues).[1][2]
|
|---|
| Cons | - −User reportsUsers describe a steep learning curve for people not used to graphs, and say tracing a mid-graph production failure means building your own observability.[1][2]
- −User reportsDurability edge cases: long tool calls (about 180s or more) silently re-execute from a checkpoint on LangGraph Cloud, and sync durability doesn't enforce write ordering.[1][2]
- −User reportsCancelling a run loses streamed state that hasn't been checkpointed yet, and the Postgres checkpointer has a long-open SSL error.[1][2]
- −User reportsDocs lag the fast-moving API, and breaking changes are a recurring complaint.[1][2]
- −User reportsSeveral teams outgrow it and write their own loop, keeping only the graph idea.[1]
| - −User reportsUsers report the whole crew hanging when one agent or API call times out; one moved the pipeline to LangGraph.[1]
- −User reportsThe early simplicity becomes a constraint: users report hard-to-reproduce drift on long tasks and a 'loose and underspecified' abstraction.[1][2]
- −User reportsForum users report poor visibility into agent decisions, roughly 10-minute crew runs, and a virtualenv near 1 GB.[1]
| - −User reportsInfra, observability and reliable scaling are left to you, and one user found those were the bigger pain.[1]
- −User reportsOne user calls it messy internally, citing inconsistent use of its own enums.[1]
- −User reportsSome teams used it extensively and then moved to their own framework.[1]
|
|---|
| Best for | Long-running, stateful, resumable workflows with human approval steps where you want explicit control flow. | Demos and internal multi-agent prototypes where speed to a first result matters more than determinism. | Python teams that want lightweight multi-agent primitives (handoffs, guardrails, tracing) with first-class OpenAI Responses and realtime support. |
|---|
| Watch out for | Make tool calls idempotent: a checkpoint replay can run them again. | Timeouts, hangs and limited observability; plan your own monitoring before production. | Production concerns like durability, scaling and approvals are on you; plan that layer early. |
|---|
| Links | | | |
|---|