AI Agents Do Not Improve by Addition: Lessons From Vercel and Agent Maintenance

Vercel improved the performance of its internal agent by removing 80 percent of its tools. The example is presented as a useful counterpoint to the common assumption that agents automatically improve when given more context, memory, tools, permissions, and autonomy.

The key, the speaker says, was careful observation of how the company's best salespeople actually worked. Rather than building around an idealized process on paper, the team studied what they ignored, what they responded to, and where human judgment still mattered, then designed the agent around the workflow that existed in reality.

Codex is strong, the speaker argues, not just because the model itself is strong, but because OpenAI keeps investing in the harness around it: terminals, desktop apps, IDEs, browsers, computer-use flows, approvals, and sandboxes.

That makes it more than a smart chatbot. It becomes a carefully maintained workbench for mechanical work. He argues that the Claude Code team is making the same strategic bet by investing heavily in the harness, treating the place where real work happens as the starting point for agent design.

The final point is that agent value has to be audited. Are people actually reading the outputs? Did the outputs change the work? After review, are they still saving time, or are they quietly adding more overhead?

Agents differ from ordinary tools because they can break in two ways at once: the surrounding world changes, and the internal models get upgraded. Maintenance is the practice of preserving fit between those two moving targets. To think about that more clearly, the speaker recommends Stewart Brand's "The Maintenance of Everything."