Harshil Agrawal is a Berlin-based senior developer educator at Cloudflare and the creator of PromptMotion, an application that generates animated videos from natural-language instructions. His work addresses a central problem in agentic software: allowing AI-generated code to execute without exposing credentials, customer data, or production infrastructure.
Originally from India, Agrawal developed his technical career through JavaScript, open-source communities, and developer education, including work as a GitHub Campus Expert and Mozilla community volunteer. At n8n, he wrote about workflow automation and built GitHub Wrap, a year-in-review application for open-source contributors that combined GitHub’s GraphQL API, n8n workflows, Cloudflare Workers, and Cloudinary. His account of its architecture documents how he assembled those services into a public-facing product.
In 2022, Agrawal joined Contentful’s developer-relations team, expanding into content modeling, TypeScript, localization, and composable web architecture. He subsequently joined Cloudflare as a developer educator, initially focusing on its data and storage products. His open-source repositories span object storage, Durable Objects, browser automation, and AI-generated applications. He also builds independent products including PromptMotion, ForwardCents, Images AI, and DocsForAI.
How he thinks about secure AI execution
- AI-generated code is untrusted code. Agrawal treats model-written programs as potentially unsafe regardless of intent. Hallucinated dependencies, runaway processes, exposed environment variables, and indirect prompt injection become serious threats when generated code inherits an application’s production privileges.
- Capability-based security. His architecture denies access by default and grants only specific, task-scoped capabilities. Cloudflare Dynamic Workers can run generated JavaScript inside V8 isolates with narrowly restricted bindings, while outbound networking remains disabled or passes through a controlled proxy. His AI Engineer Europe demonstration applies this design to an agent-generated Hacker News skill.
- Match the sandbox to the workload. Lightweight isolates suit short-lived agent tools and data transformations; Linux containers are necessary when generated applications must install dependencies, manage files, start development servers, or provide live previews. PromptMotion uses containerized execution coordinated through Durable Objects, with separate environments for individual users. Sensitive API requests pass through trusted application code, keeping credentials outside the sandbox.
- Durable workspaces should reduce real operational complexity. In his analysis of a potential PromptMotion redesign, Agrawal evaluates whether shared durable filesystems could simplify coordination between agents, previews, and rendering containers. He treats the migration as an experiment, while preserving requirements for user isolation, rendering dependencies, object storage, recovery, resource limits, execution logs, and reliable cleanup.
