APA Agent Decision Harness [](https://whop.com/aipersonaacademy) [](https://docs.typesafe.ai) [](LICENSE)

September 19, 2026 · View on GitHub

A production-grade TypeScript runtime harness for TypeSafe Jev and System One models. Powers autonomous agent decision pipelines with confidence-gated policy routing, shadow mode, trajectory verification, and automated evaluation.

Maintained and published by AIPersona Academy (APA).

🎯 Why This Exists

Autonomous coding assistants (Claude Code, Codex, Antigravity, Cursor) and multi-agent systems often stall, loop, or take unsafe actions when relying on freeform LLM chat to make control-plane decisions.

APA Agent Decision Harness wraps Jev's machine-native decisions (Choice, Score, Noul) with production control mechanisms:

  • Policy Mapping: Directly maps typed verdicts to actionable runtime verbs (execute, suppress, notify, escalate).
  • Confidence Gate: If calibrated model confidence drops below a defined threshold, fails safe or routes to human-in-the-loop review.
  • Shadow Mode: Evaluates candidate actions alongside existing agent workflows without interrupting live behavior.
  • Trajectory Verification: Inspects multi-turn agent tool calls and diffs to assert that stated goals are mathematically and semantically fulfilled.
  • Offline Evals: Replays deterministic fixtures to benchmark prompt variations with zero network latency.

📦 Installation

npm install @aipersona/agent-harness
export TYPESAFE_API_KEY=tsk_...

⚡ Quick Start

import { DecisionHarness, choice, noul, score } from "@aipersona/agent-harness";

const harness = new DecisionHarness({
  confidenceThreshold: 0.85,
  shadowMode: false
});

// Run parallel confidence-gated evaluation on live agent state
const decision = await harness.run({
  id: "agent-step-104",
  state: {
    command: "rm -rf ./dist && npm run build",
    changedFiles: ["src/index.ts", "package.json"],
    environment: "production"
  },
  questions: {
    safety: noul("Is this shell execution hazardous or destructive to production state?"),
    action: choice("What execution policy should apply?", {
      options: ["AllowImmediate", "RequireHumanConfirmation", "BlockAndLog"]
    })
  }
});

console.log("Verdict:", decision.action.choice);
console.log("Confidence:", decision.action.confidence);

🛠️ Included Recipes

  • Alert & Anomaly Filter: Suppresses noise and escalates critical infrastructure alerts.
  • Autonomous Model Router: Routes complex tasks to high-tier reasoning models and routine tasks to fast-tier models.
  • Agent Compaction & Tool Pruning: Automatically drops stale context to keep token costs minimal.

📜 License & Support