# Pranit Sharma

> Agent Harness Engineer at Vercel. Previously built Orbit at Recursive Labs.

I'm Pranit, and I make developer tools at [Vercel](https://vercel.com/). I'm currently building [fx](https://fx.sh/), a small open-source coding agent written in Zig. It runs in your terminal or inside other apps. Before Vercel I founded [Recursive Labs](https://recursive.ac/), where I built [Agents SDK](https://www.agentsdk.build/) and [Orbit](https://orbit.build/), an AI-native IDE for macOS. I also write about building software with AI.

Elsewhere: [X](https://x.com/fazxes) · [GitHub](https://github.com/fazxes) · [Email](mailto:pranit@vercel.com)

## Journey

### Agent Harness Engineer, [Vercel](https://vercel.com) (2026 to Present)

Building [fx](https://fx.sh), a tiny, open, embeddable native coding agent written in Zig.

[Website](https://fx.sh) · [GitHub](https://github.com/vercel-labs/fx) · [Docs](https://fx.sh/docs)

### Founder, [Recursive Labs](https://recursive.ac/) (2024 to 2026)

Founded Recursive Labs to build and ship AI-native developer tools, including Orbit, Agents SDK, Nightshift, OPS, X Agent, Skills, and Mosaic.

[Website](https://recursive.ac/) · [GitHub](https://github.com/Recusive)

## Projects

- [fx](https://fx.sh), 2026 - Present: A tiny, open, native coding agent written in Zig, designed as a minimal Unix-like CLI and an embeddable harness for larger systems. ([Source](https://github.com/vercel-labs/fx), [Docs](https://fx.sh/docs))
- [Orbit](https://orbit.build/), 2024 - Present: An AI-native IDE for macOS, with one agent that works across your editor, terminal, docs, and a built-in browser.
- [Agents SDK](https://www.agentsdk.build/), 2024 - Present: An open-source Rust backend for building agents on Claude, ChatGPT, and Copilot.
- [Nightshift + Recursive](https://orbit.build/nightshift), 2025 - Present: An autonomous engineering system that hardens codebases and builds features overnight, on a portable orchestration framework. ([Source](https://github.com/Recusive/Nightshift))
- [OPS](https://orbit.build/ops), 2025 - Present: A code-quality checker for AI-generated code that enforces your project's rules as files change.
- [X Agent](https://orbit.build/x-agent), 2025 - Present: An agent that drafts X replies using X's open-source Phoenix scoring model, then posts them. ([Source](https://github.com/Recusive/x-agent))
- [Skills](https://orbit.build/skills), 2025 - Present: A library of reusable agent skills for Orbit, Claude Code, Cursor, Codex, and 38+ other coding assistants. ([Source](https://github.com/Recusive/Skills))
- [Mosaic](https://www.mosaic.sh/), 2025 - Present: An in-app testing framework that runs inside your live web app and catches race conditions without mocks.

## Writings

- [High, Extra High, Max, Ultra: Reasoning Levels Are Making Models Dumber](https://www.fazxes.com/blog/reasoning-levels-are-making-models-dumber/index.md): I tested Fable 5, Opus 4.8, GPT-5.5, and GPT-5.6 at every reasoning level they ship. Past a point, more effort means more hallucination, not more intelligence. The research backs this up, and the labs aren't telling you.
- [Measure Once: Rebuilding Chat Virtualization for AI-Era Apps](https://orbit.build/blog/measure-once-ai-chat-virtualization): Our chat scrolled like it was drunk. Then we realized LLM apps break the assumptions every virtualization library was built on. Here's what we changed.
- [Introducing OPS: The System That Makes AI Write Code Like a Senior Engineer](https://orbit.build/blog/introducing-ops-orbit-protocol-server): AI writes working code that ignores your conventions. OPS is 17 subsystems of real-time enforcement that turn suggestions into laws, so every AI-written line meets your standards.
- [Introducing Nightshift: An Autonomous Overnight Agent That Actually Works](https://orbit.build/blog/introducing-nightshift-autonomous-overnight-agent): Nightshift is an open-source Python orchestrator that runs AI agents overnight to harden your codebase, build features, and produce auditable shift logs by morning.

## Skills

- [align-scope](https://www.fazxes.com/skills/align-scope/index.md): Goes through every decision in a pull request, proposal, issue, or plan with you, one at a time, before any planning starts.
- [blueprint-first](https://www.fazxes.com/skills/blueprint-first/index.md): Has an agent study the codebase before it plans or writes code, the way a senior engineer would.
- [code-simplifier](https://www.fazxes.com/skills/code-simplifier/index.md): Cleans up recently edited code for clarity without changing what it does. Making no edits is a valid outcome.
- [hand-written](https://www.fazxes.com/skills/hand-written/index.md): Rewrites prose so it reads like a skilled professional wrote it, not a language model. Has professional, semi-casual, and casual modes.
- [humanizer](https://www.fazxes.com/skills/humanizer/index.md): Rewrites AI-sounding text so it reads like the writer, without changing what it says. Based on Wikipedia's Signs of AI writing.
- [impeccable](https://www.fazxes.com/skills/impeccable/index.md): Design guidance for frontend interfaces: websites, dashboards, product UI, and native iOS and Android screens.
- [problem-clarifier](https://www.fazxes.com/skills/problem-clarifier/index.md): Interviews you about a bug or feature request and reflects it back, so the problem is clear before planning starts.
- [pure-core](https://www.fazxes.com/skills/pure-core/index.md): Keeps a pure core and pushes side effects to the edges, in any language.
- [write-swift](https://www.fazxes.com/skills/write-swift/index.md): How to write modern Swift: value types, Swift 6 concurrency, protocols and generics, API design, and testing.
- [zig-0.17](https://www.fazxes.com/skills/zig-0.17/index.md): What changed in Zig 0.17, and how to move 0.16 code over without silent behavior changes.
- [zig-best-practices](https://www.fazxes.com/skills/zig-best-practices/index.md): Guidance on correctness, ownership, error handling, and testing for Zig 0.17 and later.