The AI Agent Contract Stack

The definitive, execution-agnostic governance framework for the Agentic Economy.
A dual-layer protocol for Capabilities (UTCD) and Behavior (ABC/Contracts).

---

🏗️ The Architecture

The AI Agent Contract Stack provides a standardized way for AI systems to interact safely. It separates the description of tools from the governance of agents.

Layer 1: UTCD (Capabilities)

"What is this tool and what are its risks?" UTCD provides a declarative "Nutrition Label" for tools. It allows agents to reason about side-effects, privacy, and costs before execution.

Stored in: examples/*.utcd.yaml

Layer 2: ABC (Contracts)

"How is this agent permitted to behave?" Agent Behavior Contracts (ABC) define the enforceable boundaries—MISSION, RISK, and GOVERNANCE—that an agent must respect during its lifecycle.

Stored in: contracts/ directory

---

� Quick Start

1. Tool Capability Validation (UTCD)

Ensure your tool's "Nutrition Label" follows the standard schema:

python -m utcd.validator examples/csv-analyzer.utcd.yaml

2. Behavioral Integrity Validation (ABC)

Validate an agent's behavioral contract and verify tool risk inheritance:

python -m utcd.contract_validator contracts/examples/research-agent.contract.yaml

3. Cryptographic Signing

Secure your tool descriptors to prevent "Shadow Capability" tampering:

python demos/sign_descriptor.py examples/csv-analyzer.utcd.yaml

---

🧠 Reasoning Agent Example

The utcd package includes a built-in reasoning engine that selects tools based on your governance policy.

from utcd import UTCDAgent, Policy

# 1. Initialize with a governance policy (e.g., GDPR compliant)
agent = UTCDAgent(policy=Policy.gdpr())

# 2. Discover tools
agent.load_tools_from_directory("./examples")

# 3. Find the safest tool for the mission
results = agent.find_tools(domain="data-processing")
best_tool = results[0]

print(f"Selected: {best_tool.tool_name} (Score: {best_tool.score}/100)")

---

📁 Project Structure

Directory	Purpose
utcd/	The core Python package (Loader, Validator, Signer, RiskEngine).
contracts/	Agent Behavior Contracts (ABC) specifications and examples.
schema/	Formal JSON Schema definitions for core and profiles.
examples/	Sample UTCD tool descriptors.
demos/	Ready-to-run scenarios (Selection, Rejection, Degradation).

---

🔐 Built-in Policies

Policy	Scope	Enforcement
Policy.strict()	Safety	Rejects any tool with side-effects or data retention.
Policy.standard()	Balanced	Standard safety defaults with performance weighting.
Policy.gdpr()	Compliance	Enforces EU data residency and encryption standards.
Policy.permissive()	Audit	Provides warnings without blocking execution.

---

📖 Key Principles

1. Pre-execution only — We describe, never execute.
2. Static & Declarative — Machine-readable YAML, no runtime dependencies.
3. Offline-Valid — Guaranteed safety without network calls.
4. Immutable Trust — Canonical hashing prevents tampering after signing.
5. Agent-Centric — Designed to be ingested by LLMs for autonomous reasoning.

---

� Documentation

• WHITEPAPER.md — The philosophical and technical foundation.
• contracts/contract-spec.md — Formal ABC specification.
• FUTURE_ROADMAP.md — Our plan for an Agentic Governance ecosystem.

---

Developed by Rajan | Released under MIT License