DAgR: Distributed Agent Runtime

A serverless execution environment for autonomous AI agents.

DAgR provides a runtime where agents can schedule themselves, spawn sub-agents, and coordinate across time without keeping a process alive. It decouples agent logic from execution infrastructure, enabling truly ephemeral, distributed agent swarms.

✨ Features

• Ephemeral Agents: Agents that run, act, and exit—persisting only their state and schedule.
• Time Travel: Native scheduling. Agents can suspend themselves and wake up minutes or days later.
• Swarm Coordination: Spawn sub-agents to parallelize work across a distributed mesh.
• Serverless Native: Built on NATS JetStream for durability, exactly-once delivery, and scale.

🚀 Installation

npm install -g @addbrain/dagr

💻 Usage

The dagr CLI is your gateway to the runtime. It is designed to be used by both humans and agents (via bash tool use).

Starting an Agent

Start a local agent session (similar to Claude Code) but with DAgR superpowers enabled.

dagr start --agent claude-code

Registration (Service Discovery)

Register a specialist agent so others can find it. This uses NATS KV for service discovery.

dagr register specialist \
  --name stock-analyst \
  --description "Analyzes stock market data and Reddit sentiment" \
  --topic agents.specialist.stock-analyst \
  --config ./agents/stock-analyst/config.md

Discovery

Find available specialists in the mesh.

dagr list specialists
# Output:
# - stock-analyst: "Analyzes stock market data..."

Spawning Sub-Agents

Trigger a specialist to perform a task.

dagr spawn stock-analyst \
  --context "Check GME sentiment on Reddit" \
  --reply-to agents.main.results

Lifecycle Management

Agents can choose to suspend execution and satisfy the "Context Penalty" by exiting completely.

dagr suspend \
  --until "2026-01-22T18:00:00" \
  --resume-context "Check for stock analyst results"

The agent process terminates. The runtime schedules a wake-up event. When the time comes, a fresh container is spun up with the restored state.

🏗 Architecture

DAgR is built on NATS JetStream and Synadia Nex.

• Messaging: All agent coordination happens via NATS subjects (agents.{type}.{id}).
• State: Persistent state is stored in NATS Key-Value (KV) buckets.
• Execution: Nex (NATS Execution Engine) listens to topics and spawns workloads (containers/WASM) on demand.

Distributed Agent Runtime (DAgR) — The Distributed Harness for Autonomous Agents