How AgentField works

You write agent functions. AgentField exposes them as infrastructure: HTTP, routing, memory, tracing, policy, async, and audit.

AgentField has two moving parts:

• Control plane: routes execution, tracks workflows, manages shared memory, applies policy, records audit data, and exposes HTTP APIs.
• Agent nodes: your Python, TypeScript, or Go processes. Each node registers reasoners and skills with the control plane.

Core Primitives

The SDK gives your agent code a small set of runtime calls:

{
  "app.ai": "structured model calls",
  "app.call": "cross-agent execution",
  "app.memory": "shared state and vector search",
  "app.pause": "human approval",
  "app.harness": "dispatch harness work"
}

Request Flow

All cross-agent calls go through the control plane. That keeps routing, policy, execution history, and workflow DAGs consistent.

Sync And Async

Use synchronous execution when the caller should wait for a result:

client -> control plane -> agent -> result

Use async execution when work may run long, wait on tools, or pause for a human:

client -> control plane -> execution id
worker -> agent -> tools/agents/humans -> webhook or polling result

The important design point is that the agent stays a callable backend capability even when the work takes longer than a normal HTTP request.

Why The Control Plane Exists

Without a control plane, each agent app has to solve the same infrastructure problems:

Next: Production capabilities for the full surface, or Building Blocks for implementation details.