AgentGuard

The description states that AgentGuard is "the intent layer, human approval gate and tamper-evident ledger your agents are missing — plus an MCP server." It sits between AI agents and the real world to evaluate proposed high-risk actions before execution.

It includes:

• A policy floor with deterministic rules (caps, allowlists, thresholds)
• An intent firewall using GPT-5.6 to judge actions against declared mission
• A hash-chained ledger for tamper-evident logging
• An MCP server enabling agent integration
• A dashboard for monitoring

The product is described as evaluating actions like payments, emails, data exports, API calls, and shell commands.

Evidence The author's own write-up.

Inference This appears to be a security product for AI agents, not a general-purpose SaaS platform. It is built around the concept of "intent" rather than static rules.

The description states that AgentGuard is positioned as:

• A solution to prevent AI agent abuse (the author cites Gartner's 1 in 4 enterprise breaches from AI agent abuse by 2028)
• The "thing that watches" agents, which the author says almost nobody is building
• A product that enforces intent-based controls over agents

The claim evolution shows:

1. Initial problem: AI agents can be hijacked and act outside owner control
2. Solution: AgentGuard evaluates actions against both policy rules and declared mission intent
3. Value proposition: Prevents unauthorized actions even when all numeric rules pass

Evidence The author's own write-up.

Inference The positioning is based on a perceived market gap rather than demonstrated demand or customer validation.

The description states that AgentGuard targets:

• AI agencies (the author mentions running a one-person agency)
• Anyone who runs AI agents that send emails, touch client data, or move money through APIs
• Enterprises concerned about AI agent abuse

The author notes that "everyone is building agents" but "almost nobody is building the thing that watches them."

Evidence The author's own write-up.

Inference The target customer appears to be small to mid-sized AI agencies or enterprises with AI agent deployments, though no specific customer names or segments are mentioned.

The description does not contain any information about:

• Revenue model
• Pricing structure
• Customer acquisition costs
• Unit economics
• Monetization strategy

Evidence Not evidenced.

Inference No commercial business model is described, and the project appears to be a hackathon submission with no indication of monetization or pricing.

The description states:

• Built by one person (Matteo Misiani) using Docker, FastAPI, Next.js, OpenAI, Python, SQLite, Tailwind CSS, TypeScript
• Uses GPT-5.6 as both development tool and product component
• Includes policy floor, intent firewall, ledger, MCP server, dashboard
• 220 tests (138 backend, 82 frontend) with real artifacts required before moving to next phase
• Codex caught issues like float storage in ledger and transaction consistency
• Demo agent runs on GPT-5.6, evaluating another GPT-5.6 agent

Evidence The author's own write-up.

Inference The technical stack is self-reported and includes modern tools for backend, frontend, and AI integration. The use of Codex suggests a disciplined development approach.

The description states:

• One-person team
• Submitted to OpenAI 2026 hackathon on Devpost
• Public console available for testing
• Demo shows real-time evaluation with confidence scores and timestamps
• Tamper test functionality exists
• No revenue, customers or adoption data provided

Evidence The author's own write-up.

Inference There is no evidence of traction, revenue, or customer adoption beyond the author’s own demo. The project appears to be in early development stage with no commercial validation.

The description states:

• Gartner says 1 in 4 enterprise breaches will come from AI agent abuse by 2028
• "Everyone is building agents" but "almost nobody is building the thing that watches them"
• The EU AI Act is expected to require audit trails for autonomous systems

Evidence The author's own write-up.

Inference The competitive context is defined by a perceived market gap in AI agent security, with no mention of direct competitors or existing solutions.

Key risks and red flags from the description:

• No revenue, customers, or traction data
• Single-person team (no evidence of scaling capability)
• Product built for demo purposes only (submitted to hackathon)
• No pricing model or monetization strategy described
• GPT-5.6 used as both development tool and product component — raises questions about reproducibility and dependency risks
• No mention of compliance, legal, or regulatory readiness
• No evidence of market validation or customer feedback

Evidence The author's own write-up.

Inference The project lacks commercial viability indicators and appears to be an experimental prototype rather than a scalable business.

1. What is the actual business model? How do you plan to monetize this?
2. Have you validated demand with potential customers or partners?
3. Can you demonstrate any real-world use cases beyond your own demo?
4. How do you plan to scale from a single-person team to a commercial operation?
5. What are the risks of relying on GPT-5.6 as both development tool and product component?
6. Are there any legal or compliance implications of using AI models for decision-making in financial contexts?
7. What is your roadmap for real execution adapters, auth, and idempotency keys?

Evidence The author's own write-up.

Inference These questions are necessary to assess commercial viability and technical feasibility beyond the demo stage.