Framework architecture overview
ATFGraph vNext is a platform for building and operating multi-agent systems. Its graph server routes calls between agents and records how each run executes. You write agent nodes in Python, run them as workers, and call them through the CLI, HTTP API or language clients. Each run returns a result and an inspectable record called a path.
What the platform provides
The graph server handles work queues, retries and recovery. It also provides sessions, persistent memory, document knowledge and approval requests. Teams use these services through the SDK instead of building the storage and execution mechanisms themselves. Your application decides what to remember, which documents to ingest and when a human must approve an action.
The Console shows runs, steps, failures and costs. APIs let other applications read execution information, and optional MLflow and OpenTelemetry integrations send traces to external systems.
Agent workers call the model providers and business systems your application configures. Calls between registered capabilities go through the graph server; workers connect outbound and need no inbound port.
The three processes
| Process | What it does | Command |
|---|---|---|
| Graph server | Queues and routes work; stores paths, sessions and memory | atfgraphd |
| CLI | Manages the local server, lists agents, starts runs and reads paths | atfg |
| Worker | Runs an agent node and connects outbound to the server | python -m atfgraph.worker |
By default, a local graph server listens on 127.0.0.1:18880 and serves the Console at http://127.0.0.1:18880/ui/. The optional gRPC admin surface uses port 18882.
The server and workers must be running for agents to execute. Installing the client or publishing an endpoint does not start an agent worker.
Install and connect
If your organisation operates a managed graph server, follow the Install guide to install the client from a GitHub Release and connect to that server. The client includes the atfg CLI, Python SDK and coding-assistant skills. You do not need to build the server or Console.
For local development, follow the Quickstart. It builds the server and Console from this repository, installs the Python SDK and starts a worker. You need Rust, Node 22 and Python 3.10 or newer.
For a server you operate yourself, the Install guide also covers self-host bundles and building one with ./install.sh --source. The deployment guides cover Docker Compose, Kubernetes and Databricks.
Where to go next
| You want to | Read |
|---|---|
| Run an example locally | Quickstart |
| Understand the vocabulary | Core concepts |
| Write an agent node | Python SDK |
| Start a run and read the result | Start a traverser |
| Understand the API areas | API reference |
| Diagnose a problem | Troubleshooting |