Skip to main content

Agent Protocol Reference

Agent manifests capture LLM-backed workers in a deterministic format so you can audit prompts, safety filters, and telemetry outputs.

Core Fields​

  • agent – Name, purpose, owner, maturity stage.
  • inputs / outputs – Structured schemas with types, optionality, and pii flags.
  • guardrails – Rate limits, allowed tools, alignment policies.
  • policies – Safety, privacy, logging expectations.
  • evaluation – Test suites, judges (LLM-as-Judge), acceptance thresholds.

Usage​

import { createAgentProtocol } from '@cpms/agent';

const triageBot = createAgentProtocol(manifest);
const run = triageBot.validate();
const plan = triageBot.generateEvaluationPlan();

The evaluation helpers output mission-ready judge configurations referencing Gemini + Claude scripts from CMOS research.

Example​

{
  "protocol": "agent",
  "agent": {
    "name": "support-triage",
    "version": "1.2.0",
    "owner": "cx-automation"
  },
  "inputs": {
    "ticket": { "type": "object", "required": true },
    "history": { "type": "array" }
  },
  "outputs": {
    "resolution": { "type": "string" },
    "handoff_required": { "type": "boolean" }
  },
  "guardrails": {
    "max_tool_invocations": 5,
    "disallowed_content": ["credentials", "PII"]
  },
  "evaluation": {
    "judges": ["Claude", "Gemini"],
    "metrics": ["correctness", "tone", "policy_adherence"]
  }
}

Agent manifests make it easy to prove compliance for AI workflows without reverse engineering prompt chains.