Reference

Glossary.

The vocabulary of the Agentmatic runtime. New to graph-based agents? Start with the architecture overview or the migration guide.

StateGraph

The core graph type. A directed graph of nodes (functions) and edges (typed transitions) over a shared state object. Drop-in compatible with LangGraph's StateGraph API.

Pregel runtime

The bulk-synchronous parallel execution model behind Agentmatic. Each step is a superstep; the scheduler decides what runs in parallel within a step, and every superstep is a checkpoint boundary.

Superstep

One synchronized iteration of the Pregel runtime in which eligible nodes execute in parallel and channels are reduced. Supersteps are the unit of checkpointing and time travel.

Channel

A typed reducer that accumulates values passed between nodes. In Agentmatic, channels are implemented as lock-free SPSC ring buffers in the Rust core.

Conditional edge

An edge whose target is chosen at runtime by a routing function, letting a graph branch on state. Same add_conditional_edges API as LangGraph.

Checkpointer

The component that persists graph state at each superstep. Agentmatic ships Memory, SQLite, Postgres, Redis and S3 backends.

Time travel

Re-running a graph from any past checkpoint. Because every superstep is a checkpoint boundary, you can rewind, inspect, edit state, and resume.

Human-in-the-loop (HITL)

Pausing a graph before or after a node (interrupt_before / interrupt_after) so a person can review or edit state before execution continues.

Circuit breaker

A resilience primitive that stops calling a failing dependency after a failure threshold, then probes for recovery. Built into the open-source core.

Dead-letter queue (DLQ)

A queue that captures messages or node executions that repeatedly fail, so they can be inspected and replayed instead of lost.

Distributed runtime

A self-hosted coordinator/worker cluster that spreads graph execution across machines over gRPC — no hosted platform required.

Zero-copy state diff

Passing state between supersteps by sharing an Arc<Frame> reference rather than copying, which is a large part of Agentmatic's memory savings.

LangChain bridge

Interop helpers: from_langchain_tools() wraps any LangChain BaseTool, and as_langchain_runnable() exposes an Agentmatic agent as a LangChain Runnable.

MCP (Model Context Protocol)

An open protocol for connecting agents to external tools and data sources. Agentmatic ships an MCP client so agents can call MCP servers as tools.

Agentmatic Studio

An open-source visual debugger that mounts on a running agent: step debugger, checkpoint timeline, state-diff viewer, and graph visualization.

See also: use cases · FAQ · comparisons · Neul Labs ↗