Layers (L · I · A · T · E)

[A] Agent Harness & Loop Layer

The autonomous ReAct loop, step dispatching, and loop limits

Overview

Layer [A] defines the agent's identity, role, and autonomous reasoning loop. It is implemented in src/agent/loop.ts as ReActLoop.


The ReAct Execution Cycle

When an agent runs, it steps through an autonomous cycle:

User Prompt
    │
    ▼
[ Thought ] ──► Model analyzes context and plans next action
    │
    ▼
[ Action ] ──► Runtime dispatches tool call with parameters
    │
    ▼
[ Observation ] ──► Tool output captured and fed back to context
    │
    ▼
[ Final Answer ] ──► Structured conclusion returned to caller

Configuration in agent.json

{
  "A": {
    "name": "identity_agent",
    "version": "1.0.0",
    "intent": "Verify and validate Indian identity credentials including Aadhaar and PAN.",
    "maxSteps": 5
  }
}
  • name (required): Unique identifier for the agent class.
  • version: Semantic version string (defaults to 1.0.0).
  • intent: System directive and behavioral purpose.
  • maxSteps: Maximum number of tool iterations before returning (default: 10), preventing infinite loop runaway.
OpenLiate