Agent Preflight

Company: Agent Preflight

Self-reported basis: The description is entirely self-reported and unverified, based on a Devpost submission for the OpenAI 2026 hackathon. No external corroboration or historical data exists.

What it appears to be: A Rust-based static-analysis CLI tool that verifies agent codebases against human-approved contracts to ensure safety controls are present before deployment. It supports multiple AI agent SDKs (OpenAI, Google ADK, Anthropic) and integrates with CI/CD pipelines.

What changed: The project was submitted as a hackathon entry; no evidence of prior development or product evolution is provided.

Single most important open question: Does the tool have any real-world adoption or integration in production agent systems?

The description states that Agent Preflight is:

• A static-analysis CLI built in Rust.
• It scans agent codebases to verify that every tool decorator has the approval control its owner intended.
• It does not execute code; it parses source text using Tree-sitter.
• It uses YAML contracts for human-readable, diffable, version-controlled safety rules.
• It integrates with GitHub Actions, running 15 test suites across platforms.
• It supports OpenAI Agents SDK, Google ADK, and Anthropic Claude.

Inference: The tool appears to be a safety verification layer for AI agents, focused on ensuring that tools used by agents have the intended human approval gates before deployment.

The description states:

• The product addresses a gap in agent frameworks: “No agent framework proves, before deployment, that every high-stakes tool has the safety control its owner intended.”
• It positions itself as a static analyzer that verifies code against contracts, not a dynamic execution tool.
• It claims to support three major AI agent SDKs (OpenAI, Google ADK, Anthropic).
• The author highlights “Scan → Approve → Verify” as the workflow.
• It was built with Codex (GPT-5.6) for architecture design.

Inference: The product is positioned as a pre-deployment safety control tool, not a general-purpose agent framework or runtime. It emphasizes zero false positives, sub-second scans, and cross-platform compatibility.

The description states:

• The tool targets agent developers who use OpenAI Agents SDK, Google ADK, or Anthropic Claude.
• It is aimed at those who want to ensure that tools used by agents have intended human approval controls.
• It is designed for CI/CD pipelines, suggesting a developer or DevOps audience.

Inference: The ICP appears to be AI agent developers and teams building autonomous systems, particularly in environments where safety and control are critical.

The description states:

• No pricing, business model, or monetization strategy is mentioned.
• It is a CLI tool built for open-source or internal use.
• It is presented as a hackathon submission, not a commercial product.

Inference: There is no evidence of a business model or pricing structure. The tool appears to be self-hosted and open-source in nature, with no indication of paid features or subscriptions.

The description states:

• Built in Rust — zero-dependency, single static binary.
• Uses Tree-sitter for parsing Python ASTs directly from source text.
• YAML contracts are used for human-readable safety rules.
• GitHub Actions integration with 15 test suites across Windows, macOS, and Linux.
• The tool is designed to be conservative, failing if it cannot prove a control exists.

Inference: The technical approach is static analysis, not dynamic execution. It is built for cross-platform compatibility and CI/CD integration, with an emphasis on safety over false positives.

The description states:

• Submitted to the OpenAI 2026 hackathon.
• The author mentions 15 CI test suites across platforms.
• It is a single-person project (Pratik Yadav).
• No evidence of customers, revenue, or adoption is provided.

Inference: There is no traction or maturity evidence. It is a hackathon prototype, not a product in production use.

The description states:

• The tool addresses a gap in agent frameworks.
• It supports OpenAI Agents SDK, Google ADK, and Anthropic Claude.
• No direct competitors are named.

Inference: The competitive context is AI agent safety tools, but no evidence of existing or competing solutions is provided. The tool appears to be novel in its approach (static analysis vs. dynamic execution) but lacks a clear market positioning or competitive landscape.

The description states:

• It is a single-person project.
• It was submitted as a hackathon entry, suggesting early-stage development.
• No evidence of real-world adoption, CI/CD integration, or customer feedback.
• The tool is conservative in design, failing when uncertain — which may be a feature but also limits flexibility.

Inference: Key risks include:

• No product-market fit demonstrated.
• Limited team capacity for scaling or support.
• No evidence of real-world use or integration.
• High-risk conservatism may limit usability in complex agent systems.

1. What is the actual use case or problem you're solving in production?
2. Have any teams or organizations adopted this tool in their CI/CD pipelines?
3. How does it handle edge cases or complex tool decorators that may not map cleanly to YAML contracts?
4. Is there a plan for monetization or commercialization beyond the hackathon?
5. What are the limitations of static analysis in agent safety, and how do you address them?