AgentGuard CI

Inspiration

AI agents are moving from chat into real enterprise action: they can repair code, trigger browser/RPA workflows, query data, update records, write customer responses, and coordinate with other agents. The risk is that an agent can look successful while leaking data, skipping approval, mutating production, weakening tests, or inventing evidence.

AgentGuard CI was built for UiPath AgentHack Track 3: UiPath Test Cloud. The core question is: before an AI agent is allowed into a real workflow, can we test whether its action is evidence-backed, bounded, reversible, and routed to a human when needed?

What it does

AgentGuard CI is a reliability firewall for enterprise AI agents. It runs repeatable agent failure scenarios and classifies each result as:

• Promote: evidence-backed and safe enough to continue.
• Review: useful, but requires a human decision.
• Block: unsafe, overconfident, destructive, missing approval, or missing evidence.

The prototype includes a command-backed code-repair suite, a live-local enterprise agent adapter suite across browser/RPA, data, support, workflow, document, finance, HR, CRM, SOC, knowledge, and multi-agent scenarios, plus JSON, Markdown, JUnit, and Test Cloud-style evidence packets.

How UiPath is used

The project is designed around UiPath Test Cloud as the governance layer for agent reliability testing. AgentGuard maps each failure mode into a repeatable Test Cloud-style case, attaches evidence artifacts, and produces a clear promote/review/block decision that a human reviewer can inspect before a release or workflow action is trusted.

The UiPath Labs environment hosts the AgentGuard review-agent path and demonstrates the final review gate: AgentGuard evidence goes into UiPath, UiPath reviews the evidence, and the output is a release decision with risk score, rationale, and required follow-up.

How I built it

The system is a TypeScript monorepo with a React/Vite dashboard, Node.js scenario runners, and reliability-core packages. The architecture separates three layers:

• Scenario runners execute repeatable agent failure modes.
• Reliability gates normalize results into goal fidelity, tool boundary, evidence integrity, state safety, and human approval checks.
• The dashboard and Test Cloud evidence model make every decision inspectable by judges, developers, and reviewers.

Challenges I ran into

The hardest part was making the product honest. It would be easy to claim broad agent support without evidence. Instead, AgentGuard separates command-backed tests, live-local deterministic traces, and UiPath-hosted review evidence so judges can see exactly what was tested.

Another challenge was making the UiPath integration visible. The current submission shows the integration as an evidence review gate: AgentGuard produces evidence, UiPath reviews it, and Test Cloud-style cases make the result repeatable.

Accomplishments that I'm proud of

• A working AI-agent reliability gate with repeatable local verification.
• Test Cloud-style evidence for code repair and enterprise agent scenarios.
• Promote/review/block outcomes instead of vague pass/fail summaries.
• A public GitHub repository with setup instructions, MIT license, deck, and evidence artifacts.
• A UiPath Labs review-agent path that connects AgentGuard evidence to a hosted review decision.

What I learned

Agent testing is not just about whether an agent completed a task. It is about whether the action is safe enough to approve. AgentGuard turns reliability, security, and human approval into a repeatable testing workflow.

What's next

Next, I would connect the evidence packet directly into live UiPath Test Cloud cases, add richer Studio Web orchestration, and benchmark multiple coding and low-code agents under the same reliability contract.

Built With

• agent-reliability
• ai-agents
• automation
• ci
• coded-agents
• enterprise-ai
• github-actions
• governance
• node.js
• playwright
• react
• studio-web
• typescript
• uipath
• uipath-test-cloud
• vite
• vitest