PREVIEW

Note: This analysis is based entirely on the author’s own account. No external verification or historical data exists for this project.

The description states that PREVIEW is:

• A transactional runtime between AI agents and tools they can change.
• An isolated copy of the world where agents stage proposed actions.
• A system that turns those staged actions into an immutable cross-system effect graph.
• A tool that allows users to:
  • Approve the exact staged future.
  • Add constraints and replay from clean shadow state.
  • Abort transactions with zero provider writes.

It is described as a provider-neutral TypeScript modular monolith, with a separate durable worker. The system includes:

• Shadow state maintained with cross-adapter read-your-writes.
• Immutable effects sealed into deterministic graph digest.
• Separation of simulation and commit capabilities.
• Integration with GPT-5.6 via Responses API, but without agent-level approval or commit rights.
• Commit worker that revalidates observed read sets, executes in dependency-aware order, and verifies outcomes.

Inference: The system is built for safety and traceability rather than general-purpose automation.

The author claims:

• AI agents are dangerous not just when they fail, but when they succeed at doing the wrong thing.
• Most agent safety systems inspect permissions or prompts, but real danger emerges only after multiple tool calls combine across systems.
• PREVIEW enables developers to preview deployments before going live — analogously, it allows users to preview the future an agent is about to create.

The positioning evolves from:

1. A conceptual problem (agent safety).
2. To a technical solution (a runtime with shadow state and effect graphs).
3. To a practical framework (SDK or MCP-compatible gateway for certified adapters).

Claim: PREVIEW is positioned as a boundary layer for AI agents to prevent unintended consequences.

The description does not name specific target customers, but implies:

• Software developers who build or manage AI agents.
• Teams using AI agents in production, particularly those working with tools like CRM, email, calendar, payments, and infrastructure.
• Organizations concerned with agent safety and compliance, especially where human oversight is required.

The author notes that PREVIEW is designed to become a horizontal transaction boundary for any system where agents can create consequential change.

Inference: The ICP likely includes enterprise or developer teams working in regulated or high-risk environments, where AI agent actions must be auditable and reversible.

Not evidenced.

The description does not mention:

• Revenue model.
• Pricing strategy.
• Monetization approach.
• Customer acquisition or retention plans.

Note: No business model or pricing data is present in the self-reported account.

The author states:

• PREVIEW is built as a TypeScript modular monolith with separate durable worker.
• Uses GitHub, Gmail API, Google Calendar API, OpenAI API, Node.js, OAuth 2.0, PostgreSQL, Vitest, Zod, etc.
• Implements deterministic adapters, shadow state, and read-your-writes.
• Separates simulation from commit.
• Includes replay invalidation, crash recovery, and journaling execution intent.
• Has a Judge Mode for testing without credentials.
• Demonstrates integration with real systems (e.g., Google Calendar move, Gmail send).

Inference: The system is technically sophisticated and built with safety and traceability as core principles.

Not evidenced.

The description does not include:

• Revenue or ARR.
• Customer base or adoption metrics.
• Product usage data.
• Market traction or user feedback.
• Any evidence of product-market fit beyond the hackathon demo.

Note: The project is described as a single-person hackathon submission with no external validation or commercial deployment.

Not evidenced.

The description does not:

• Name competitors.
• Describe competitive advantages.
• Compare to existing agent safety tools or transactional systems.

Inference: Given the focus on agent safety and cross-system simulation, it may relate to areas like workflow automation, compliance platforms, or AI governance tools — but no direct comparison is made.

1. Single-person project: The entire system was built by one developer (Aru Chauhan), which raises questions about scalability, long-term maintenance, and team capacity.
2. No commercial traction: No evidence of customers, revenue, or product-market fit beyond a hackathon demo.
3. Unproven adoption: The author states that PREVIEW is not yet deployed in production environments — only demonstrated in sandboxed mode.
4. Limited scope: The current implementation focuses on mail and calendar; the next steps involve building an SDK or gateway — suggesting it's still early-stage.
5. High technical complexity: While impressive, the system requires deep integration with many providers and complex logic for replay, reconciliation, and crash recovery.

Red Flag: The lack of any commercial or user data makes it difficult to assess viability or scalability beyond a prototype.

1. What is your plan for scaling beyond the current single-developer architecture?
2. Have you identified potential enterprise use cases or partners who might adopt this system?
3. How do you intend to monetize or commercialize PREVIEW?
4. Are there any known limitations in how well it handles complex integrations or edge cases?
5. What are the key assumptions behind your vision for a horizontal transaction boundary for AI agents?
6. Can you describe how you would onboard and train users on the approval and replay workflows?