gMAS and LangGraph

gMAS and LangGraph both execute graph-shaped LLM workflows. LangGraph provides runtime routing, commands, interrupts, checkpointing, and graph migrations. gMAS exposes the role graph and remaining plan to a bounded set of edits during the run.

The table compares public APIs rather than behavior an application could build with arbitrary Python. LangGraph details were checked against its documentation on 22 July 2026.

Design comparison

Concern	LangGraph	gMAS
Primary abstraction	A compiled state graph whose nodes update shared state	A RoleGraph containing agents, task nodes, communication edges, local state, and graph features
Runtime routing	Conditional edges and Command(goto=...) select among authored destinations	Conditions select routes; TopologyAction can also skip/force agents, add/remove edges, insert recovery chains, or change the final responder
Topology lifecycle	Graph definitions can be migrated; interrupted threads have documented constraints on removing or renaming nodes	The same graph object and remaining plan can be edited during an active run
Persistence and replay	Built-in checkpoints, state history, time travel, interrupts, and durable resume	Agent state, optional storage, callbacks, streaming events, and JSONL traces; it is not a replacement for LangGraph's durable checkpoint stack
Parallel work	Parallel supersteps and Send/routing patterns	Parallel execution groups computed from the live role graph
Graph-learning data	Application-defined state and graph inspection	Adjacency tensors, edge attributes, embeddings, and PyTorch Geometric export are part of the graph API
Typical fit	Durable stateful workflows, human-in-the-loop execution, and deployments built around LangGraph's runtime	Experiments or systems where role topology, routing features, and bounded same-run edits are the object being controlled

Bounded edits in gMAS

A topology hook receives StepContext after a step and may return TopologyAction. The runner validates the action before applying it to the graph or remaining plan. Actions include:

• early stop with a recorded reason;
• skip or force selected agents;
• add or remove communication edges;
• insert recovery work;
• change the final agent;
• rebuild the remaining plan when required.

The finite action set can be budgeted, tested, traced, or rejected by the runner. Hooks do not execute unrestricted graph mutation through this interface.

LangGraph capabilities

LangGraph is not a fixed sequential graph. Its public API includes:

• conditional edges and Command for runtime control flow;
• checkpointed persistence and state history, including replay, forks, and pending-write recovery;
• interrupts for durable human-in-the-loop execution;
• documented graph migrations, including topology changes with constraints for interrupted threads.

The gMAS-specific surface is the combination of mutable role topology, graph-learning features, scheduler integration, and result metrics in one runtime model.

Practical choice

The choice depends on which runtime behavior the application needs:

• Prefer LangGraph when durable checkpointing, time travel, deployment infrastructure, or an established state-graph ecosystem is the primary requirement.
• Prefer gMAS when the experiment or application needs to inspect and modify role topology during the run, attach model features to that topology, or evaluate routing and pruning policies directly against graph state.
• Use the matched topology benchmark when comparing operational behavior. It records prompts, models, datasets, concurrency, and framework-specific results instead of inferring performance from API design.

Neither framework makes additional agents free. The number of model calls, the topology, the tool budget, and stopping policy still determine most of the operational cost.