September 14, 2026
AI products now sell work rather than software: Jim finds candidates itself
On September 13, 2026, Peter Yang showed Jim, Brex’s AI recruiter: it searches for candidates, filters applications, and sends emails. According to VentureBeat’s review from August 10, 2026, Jim is built on OpenClaw and works with Greenhouse and LinkedIn. Brex is selling not a screen for recruiters, but ready-to-hire candidates.
Peter Yang
@petergyang
The bar for building good AI products is much higher now because you can sell the work itself, not software. From @pedroh96, CEO of Brex: “The reality is that all good AI products are structured the same way. It is an agent loop and a set of tools. You give models access to tools, start the agent loop, and let it do its job.” Here is one example Pedro shared in our interview: Instead of making recruiters use software to find candidates, Brex built Jim, an AI recruiter that uses Greenhouse, LinkedIn, and other software to automatically identify suitable candidates. This fits my view that Human ↔ software is rapidly becoming Human ↔ agent ↔ software. 📌 Full episode here: https://youtu.be/LE0LNULrsEM
"Stop building AI agents. Build AI employees instead." Here’s my new episode with @pedroh96 (co-founder @brexhq), who showed me why AI employees with a specific job, skills, manager, and budget are better than AI agents, including: → A live demo of Jim, his AI employee to source and recruit job candidates → How to prevent your AI agents and employees from leaking company data → Why he still uses OpenClaw to run his life and $5B business Some quotes from Pedro: “Focus on building virtual employees versus open-ended agents that could do a million things.” “You’re now able to sell the work itself versus just selling the software.” “Because it’s so easy to build things now, having discernment on what matters is more important than ever. What is a true lever to actually deliver value in a unique way? Let’s go do this one thing, even though we could do five or ten." 📌 Watch now: https://youtu.be/LE0LNULrsEM Thanks to our sponsors: @RiversidedotFM: All-in-one AI studio for podcasts and video https://creators.riverside.com/PeterYang @linear: The AI agent platform for modern teams https://linear.app/behind-the-craft
· 15.2K views
Recruiters used to use software themselves to find candidates. Now Brex assigns this part of the job to Jim and leaves ambiguous cases to humans.
The framework. In Pedro’s formulation, a good AI product has an agent loop and a set of tools. A trigger starts a run, an external check decides “done” or “try again,” and a stop condition closes the loop.
Control. In April 2026, Brex described CrabTrap: the proxy first checks a URL against static rules, then sends all other requests to an LLM judge. In one production workflow, the judge triggered on fewer than 3% of requests, while an out-of-policy email prompted manager review in Slack. Quick Start launches the proxy with `docker compose up -d`, after which the agent connects through `HTTP_PROXY` and `HTTPS_PROXY`. HTTPS requires a generated CA certificate in the runtime environment or system.
The trajectory Peter Yang describes as Human ↔ software and Human ↔ agent ↔ software leads to AI employees with specific jobs, rather than general-purpose agents.