Distributed

Cluster agents without leaving your VPC.

Agentmatic's distributed runtime is self-hosted, gRPC-based, and built into the engine. Coordinator/worker for centralized control, or P2P mesh for resilience. No SaaS dependency.

Coordinator/worker setup

from agentmatic.cluster import ClusterConfig, Worker

# On each worker box:
Worker.serve(addr="0.0.0.0:9090")

# In your application:
config = ClusterConfig(
    topology="coordinator-worker",
    transport="grpc",
    workers=["worker-1:9090", "worker-2:9090", "worker-3:9090"],
    load_balancing="least-loaded",
)
agent = Agent.builder("distributed").llm(OpenAI()).cluster(config).build()

Peer-to-peer mesh

config = ClusterConfig(
    topology="p2p",
    transport="grpc",
    peers=["peer-1:9090", "peer-2:9090", "peer-3:9090"],
    consensus="raft",
)
agent = Agent.builder("p2p-distributed").llm(OpenAI()).cluster(config).build()

When to cluster

• High concurrency. Thousands of simultaneous agent runs — single-host event loops queue, clusters parallelize.
• Long-horizon workflows. Multi-hour graphs that need checkpoint hand-off if a worker dies.
• Map fan-out. Doc pipelines that fan to 100s of LLM calls — distribute across CPU.
• Multi-region. Run workers in geo-distributed nodes to meet data-residency requirements.

What it doesn't do

Agentmatic clusters are not a replacement for Kubernetes / your job scheduler. The cluster ships gRPC + checkpoint hand-off; you still run it under k8s, systemd, or your platform of choice. We don't hide the runtime from you.

Observability

• OTLP traces emitted by every worker, correlated by graph run ID.
• /healthz on every worker for k8s probes.
• agentmatic.cluster.stats() returns per-node throughput, queue depth, p95 latency.