AgentGuard

AgentGuard uses AI agents, coding agents, and enterprise orchestration to evaluate, repair, certify, and govern AI systems before deployment.

Comment

Inspiration

As AI agents get smarter, companies are quickly adding copilots, internal assistants, and automated workflows into daily business. But normal software testing tools cannot answer the big security questions that companies worry about:

  • Can someone trick or break this agent?
  • Will it accidentally leak sensitive customer data?
  • Does it actually follow company rules and policies?
  • Did our latest prompt update introduce hidden bugs?
  • Is this system really safe to run?

I noticed a massive gap today: almost every company has strict approval rules for normal software releases, but almost nobody has a clear release process for AI agents.

That inspired me to build AgentGuard : an Enterprise AI Release Platform that acts as the final security and safety approval gate before AI agents go live.

What it does

AgentGuard tests AI agents through a strict, multi-step safety check before they are approved for release.

Whenever a team submits an agent release request, my platform automatically:

  • Reads project requirements and safety rules
  • Creates custom test cases and difficult edge cases
  • Runs automated security attacks to test for prompt injection and hacking
  • Checks for data leaks and policy violations
  • Creates automatic code fixes and prompt improvements when checks fail
  • Calculates trust and safety scores
  • Issues digital safety certificates or sends borderline cases to a human boss for review

Instead of testing an AI agent just once, AgentGuard tracks the full history of an agent across different versions.

For example:

  • Version 1.0 might be blocked because it failed security tests.
  • Version 1.1 might pass security checks but trigger a rule warning that generates an automatic code fix.
  • Version 1.2 passes all safety checks, earns digital certification, and automatically starts running in production.

How I built it

I built AgentGuard as a full-stack platform designed for speed and reliability.

Frontend

Built with Next.js 14, TypeScript, Tailwind CSS, and Lucide Icons. The web dashboard gives engineering and safety teams:

  • An Executive Board showing live project statuses
  • A History Dashboard tracking agent versions over time
  • Deep analytics and test logs
  • A Settings Center to adjust safety rules

Backend Core

Powered by Python 3.11+, FastAPI, Pydantic, and SQLAlchemy backed by SQLite. The backend manages project workflows, tracks scores, and keeps a permanent log of all safety checks.

The 7-Agent Safety Pipeline

To test complex AI systems thoroughly, I built a team of 7 specialized Python testing agents:

  1. Requirement Agent: Checks if the agent matches business goals.
  2. Test Generation Agent: Creates automated boundary tests.
  3. Adversarial Agent: Runs security attacks to test for hacking vulnerabilities.
  4. Execution Agent: Runs agent code inside safe test environments.
  5. Failure Analysis Agent: Simulates network timeouts and system errors.
  6. Repair Agent: Suggests exact code fixes and prompt changes when tests fail.
  7. Governance Agent: Combines all test scores into a final pass or fail decision.

A smart AI connection supports live AI models while keeping backup rules so tests stay fast, consistent, and reliable during live demos.

UiPath Ecosystem Integration

I built AgentGuard to connect directly into the UiPath enterprise automation suite:

  • UiPath Orchestrator Bridge: When an agent gets a passing safety score, my UiPathMaestroAdapter makes REST API calls directly to Orchestrator (StartJobs), automatically starting production workflows (HTTP 201 Created).
  • UiPath Action Center Approval Gate: When an agent's safety score falls into a borderline range, my UiPathActionCenterAdapter queries UiPath Action Center via REST API, creating a human task where a compliance officer can review test reports and approve or block the release.
  • Flexible Architecture: I built interface adapters (UiPathTestCloudAdapter, UiPathIntegrationServiceAdapter) structured for local test running and event logging, ready to connect directly into enterprise grids.

Challenges I ran into

One of my biggest challenges was designing a platform that felt like a real safety approval system rather than just another AI chat dashboard. I focused heavily on tracking scores, version history, and release gates instead of just showing raw logs.

Another big technical challenge was dealing with unpredictable AI responses while keeping enterprise cloud connections stable. Sending live REST API calls to UiPath Orchestrator and Action Center required careful handling of security tokens and cloud permissions. I solved this by verifying my compiled automation packages natively on Windows Desktop runtimes while keeping clean REST API bridges.

Accomplishments that I'm proud of

  • Designing and building a complete AI release safety process from scratch
  • Building a collaborative 7-agent testing and security pipeline
  • Connecting my backend directly to UiPath Orchestrator via live REST API (StartJobs) to trigger real automation jobs automatically after approval
  • Integrating UiPath Action Center task queries to show live human approval controls
  • Generating automatic code fixes and prompt updates via my Repair Agent
  • Creating a clean, executive-ready Next.js dashboard that visualizes technical risks in clear business terms

What I learned

I learned that testing AI systems is very different from testing regular software. Passing unit tests isn't enough; enterprise AI needs continuous proof of safety, accuracy, and operational trust.

I also learned how powerful it is to pair AI testing pipelines with strong enterprise automation platforms like UiPath. By connecting security testing directly to Orchestrator and Action Center, I showed how companies can automate safety approvals without removing human supervision.

What's next for AgentGuard

  • Connecting my test adapters into live cloud webhooks for UiPath Test Cloud and UiPath Integration Service
  • Building live production monitoring that watches agents while they work and revokes safety certificates if behavior drifts
  • Creating an automated Git tool that directly submits code fixes suggested by my Repair Agent
  • Building an enterprise marketplace of pre-made safety policy packs (HIPAA, GDPR, EU AI Act)

I believe the future of AI requires the same strong safety standards that exist for regular software releases today, and AgentGuard is my step toward making enterprise AI safe and reliable.

Built With

Share this project:

Updates