@4shubGitHub ↗

Shubham Naik

Engineer at Letta coordinating persistent cloud coding agents through chat.letta.com, with Zed and connected engineering tools for implementation, debugging, and review.

How this setup has evolved

Version 1 · Latest

Shared this setup with the community.

Shubham Naik’s AI setup

What I build

I’m an engineer at Letta working on coding-agent reliability and developer workflows. I use AI for implementation, debugging, CI diagnosis, and turning workflow failures into product fixes that help future agents.

My everyday setup

  • Letta Cloud on chat.letta.com (current, primary): Most of my work now happens with persistent agents in cloud sandboxes. I can run investigations and implementation in parallel without tying up my Mac’s disk or CPU.
  • iMessage (current): I can text my agent from my phone to start work, follow up, or keep a task moving away from my desk.
  • Letta Desktop (current, local): I use it when an agent needs my Mac, local tools, or a desktop interaction. The same conversation can move between cloud and local environments.
  • Zed (current): The editor I’ve started using for navigating the codebase, inspecting changes, and making hands-on edits.

Connections & context

  • GitHub (connected): Gives agents repository, pull request, review, and CI context.
  • Linear (connected): Keeps implementation tied to engineering tickets and active work.
  • Slack (connected): Supports team communication and follow-up around ongoing work.

MCP connections

  • Better Stack (connected): Lets my agent investigate production logs, traces, metrics, and service health.
  • LaunchDarkly (connected): Gives my agent context and tools for working with feature flags.
  • Sentry (connected): Supports error and issue investigation from the same agent workflow.

How I delegate and review

  • I coordinate multiple agents from chat.letta.com and keep independent work moving in parallel while earlier changes are in CI.
  • I prefer small, reviewable batches—usually around four or five files. Once a requested feature is coherent and validated, my agent can create the pull request without waiting for another instruction.
  • If an agent workflow fails, I investigate the product-level cause and improve the repository so other agents avoid the same failure.
  • I expect the full relevant CI set and focused validation, not just a plausible diff. For UI work, I want visual proof before merge. Merging remains the human checkpoint, even when CI is green.

What runs where

  • Letta Cloud (current, primary): Runs isolated agents, parallel investigations, implementation, and longer-running validation without consuming local resources.
  • MacBook (current, local): Runs Letta Desktop, Zed, and browser-based flows that need my direct interaction. I move an active conversation here when local tools or authentication are needed.