---
title: "The Builder"
slug: may-2026
date: 2026-05-10
description: "Three bets are forming in dev tooling: the harness, the integration layer, the app. Plus the Cursor SDK, Warp open source, Claude Design, and MCP."
edition: 3
canonical: https://lukestahl.io/newsletter/editions/may-2026/
---

Three bets are forming in the developer tooling space. The harness is the product (Pi, Claude Code, Cursor). The integration layer is the product (Executor, Composio). The app itself is the product (Agent Native). The protocol question matters too. Perplexity wants direct integrations. Anthropic bets MCP becomes the standard. The thing worth paying attention to: whoever controls the harness controls the memory. That's where lock-in actually lives. Unrelated: has anyone noticed every AI company's logo looks like a butthole now?

## Blogs & Projects

- <img src="/blog/images/stop-upgrading-your-model-fix-your-harness/Harness-agent-hero_png.png" alt="Stop upgrading your model. Fix your harness." class="li-thumb"><span class="li-text">**[Stop upgrading your model. Fix your harness.](https://lukestahl.io/blog/stop-upgrading-your-model-fix-your-harness/)**<br>The harness around the model matters more than the model itself. Context management, memory, tools, guardrails. Get those right and model selection stops being the conversation.</span>
- <img src="/blog/images/building-dev-tools-for-agents/humans-and-agents_png.png" alt="Building developer tools for agents (not just humans)" class="li-thumb"><span class="li-text">**[Building developer tools for agents (not just humans)](https://lukestahl.io/blog/building-dev-tools-for-agents/)**<br>AI agents are the new consumer of your API. Claude Code, Copilot, and others need machine-readable responses, consistent schemas, and semantic error codes to work autonomously. If your API wasn't built for agents to call, it's going to get skipped.</span>
- <img src="/blog/images/plg-for-developer-companies/plg_png.png" alt="PLG for developer companies" class="li-thumb"><span class="li-text">**[PLG for developer companies](https://lukestahl.io/blog/plg-for-developer-companies/)**<br>PLG in developer companies comes down to the free tier being your acquisition channel. Let developers build something real before asking them to pay. Time-based trials work against adoption. Stripe and Vercel are the models worth studying.</span>

## Links I like

- **[Where the goblins came from](https://openai.com/index/where-the-goblins-came-from/)**<br>A GPT-5 personality preset rewarded creature-heavy language during training. The result: goblin metaphors everywhere. Small training incentives compound into visible behavior shifts. Worth understanding if you're evaluating model providers or building anything on top of RLHF-tuned models.
- **[Build programmatic agents with the Cursor SDK](https://cursor.com/blog/typescript-sdk)**<br>Cursor shipped a TypeScript SDK for building coding agents. You write the agent logic, they handle the infrastructure. Runs on their cloud VMs with sandboxing and session management. Plugs into CI/CD. The pitch is you skip maintaining your own agent stack.<br>_Roshan Sadanani_
- **[Warp is now open-source](https://www.warp.dev/blog/warp-is-now-open-source)**<br>Warp open-sourced its client and runs contributions through Oz, a cloud agent that handles implementation. Contributors focus on product decisions and verification. The agent does the coding.<br>_Zach Lloyd_
- **[100 B2B Startups, 100+ Stats, and 14 Graphs on Web, Social, and Content](https://newsletter.mkt1.co/p/state-of-marketing-report-web-social-content-part-2)**<br>Analysis of 100 B2B startups across web, social, and content. Real data, not survey results. Useful for benchmarking your own marketing against what's actually happening.<br>_Emily Kramer_
- **[Introducing Claude Design](https://www.anthropic.com/news/claude-design-anthropic-labs)**<br>Claude Design lets you generate and iterate on visual work inside Claude. Prototypes, marketing assets, presentations. It pulls in brand elements automatically and supports real-time edits. Runs on Opus 4.7.
- **[Building agents that reach production systems with MCP](https://claude.com/blog/building-agents-that-reach-production-systems-with-mcp)**<br>Three ways to connect agents to external systems: direct API calls, CLIs, and MCP. MCP wins for production because it standardizes the connection. One protocol, scalable, and already adopted across major platforms.

## Technology was a mistake

- [Why do AI company logos look like buttholes?](https://velvetshark.com/ai-company-logos-that-look-like-buttholes)<br>_Radek Sienkiewicz_
- [I Will Never Respect A Website](https://www.wheresyoured.at/i-will-never-respect-a-website/)<br>_Ed Zitron_

## Technical & Developer Marketing Jobs

- [Developer Marketing at Linear](https://linear.app/careers/62a4524c-75da-499e-b150-7f68ce9f2da3)
- [Product Marketing Manager at Cursor](https://cursor.com/careers/product-marketing-manager#apply)

## Folks to follow

- [Rhys Sullivan](https://x.com/RhysSullivan)
- [Thariq](https://x.com/trq212)
- [James Sun](https://x.com/JamesZmSun)

---

That's it for May. If you're rethinking how your team talks to developers, I'm probably thinking about the same stuff. Find me on [LinkedIn](https://www.linkedin.com/in/lukestahl/) or [X](https://x.com/LukeStahl11).

— Luke