Video summary
My Agentic Development Stack
Main summary
Key takeaways
Summary
Ras Mic walks through his current agentic development stack: the coding environments, models, frameworks, services, and automation he uses to build software. His central idea is to use whichever tool best fits a task rather than commit to one provider.
Coding Environments and Agents
- T3 Code is his usual development hub. He values its ability to use different tools’ native harnesses, connect to machines across his network—including Linux and Mac—and distribute agents across them. Its project-and-thread sidebar also helps him track work like an inbox, with tasks marked complete when their pull requests are merged.
- OpenAI Codex is especially useful to him for computer-based work, research, and hardware or 3D projects. He describes using it with Blender to plan a physical project and for practical tasks such as budgeting and improving his workspace.
- Claude Code is primarily a CLI tool in his workflow, particularly when working over SSH or configuring a remote machine. He says he rarely uses its graphical app.
- Cursor and Devin are mainly used through graphical or web interfaces for cloud-agent work. He connects agents to Linear through MCP, then assigns tasks—such as production bug fixes—so a pull request can be prepared while he investigates the issue.
- He recommends comparing subscriptions and switching to whichever currently offers the best value, rather than staying loyal to a single provider.
Models
He says his main coding model is Claude Opus 5.5, praising its balance, ability to follow instructions, and performance on longer tasks. He considers GPT-6 Astra useful for research and deeper codebase analysis, but says he currently uses Opus more. These model comparisons reflect his assessment at the time of recording.
Application Stack
- SvelteKit/Svelte is his preferred web framework, with Next.js also in use. He favors Svelte for its AI compatibility and the fast, responsive applications he says it produces by default.
- Convex is a frequent backend choice because of its type safety, real-time behavior, and integration with Svelte. Its components let agents set up services such as Stripe, Gmail, PostHog, Browserbase, and WorkOS using provided documentation or links.
- Effect is his standard for backend services and external-service code. He says its structured error handling helps both him and coding agents build more robust software.
- PlanetScale is his Postgres provider when he does not use Convex.
- For deployment, he uses Vercel or Cloudflare. When choosing Cloudflare, he tends to use its broader stack, including Workers, R2, Hyperdrive, and Durable Objects.
Tools and Integrations
- Greptile reviews pull requests and assigns a confidence score. His “grep loop” skill feeds review feedback back to the coding agent until the score reaches 5/5.
- Paper is presented as an AI-friendly design canvas and Figma alternative, with an MCP integration that lets agents create designs. He says he used it for portfolio and product redesign work.
- Mobbin is a source of website and app design inspiration and also offers an MCP server.
- Whisper Flow is his voice-to-text tool for dictating instructions to agents.
- Linear tracks team issues and work status; its MCP integration allows cloud agents such as Cursor or Devin to be assigned tasks.
- Orgo provides a separate computer for an agent. He also mentions Oracle in connection with computers used by his Pluto project, while noting that his investment in the company could affect his perspective.
Walkthroughs and Workflow Guide
The video includes a sponsor demonstration of GrokBot. The presenter shows separate bots for engineering, design, and newborn-care questions. In the design workflow, he sends a website screenshot to a designer bot, has it create a design in Paper, then starts a group chat with an engineer bot to implement the design and return a deployed URL. He also demonstrates using GrokBot on a phone to coordinate six Cursor Cloud Agent tasks, resolve merge conflicts, and merge the resulting changes.
He also explains his open-source “software factory” workflow:
- The agent reads his skills repository and
agents.mdinstructions. - It creates an isolated worktree so multiple agents can work in parallel without conflicting with the main branch.
- It writes code using his code-structure skill.
- It supplies evidence that the change works, such as a video recording for UI changes or tests. If recording is unavailable, it adds before-and-after screenshots to the pull request.
- Greptile reviews the pull request. If the score is below 5/5, the grep-loop skill sends the feedback back to the coding agent for fixes.
- The result is a pull request with a 5/5 score for him to review and merge.
Main Speaker and Sources
- Main speaker: Ras Mic, who also identifies himself as Ross in the closing.
- Sponsored segment: GrokBot, identified in the video description as a SpaceXAI sponsor.
- Other sources discussed: The presenter’s own experience with the named tools, models, services, and open-source skills.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.