AI News & Strategy Daily | Nate B JonesJune 17, 202618m

Don't build more AI agents until you watch this

TL;DR

Vercel's agent got better by removing tools, not adding them: After studying a top sales rep's real inbox workflow, Vercel found the agent became more trustworthy when the team pruned its toolset instead of piling on more skills, memory, and integrations.
The real product is the harness around the model: Nate frames the agent as the worker and the harness, or workbench, as everything around it: files, tools, memory, approvals, proof requirements, boundaries, and human review.
Agents can break when models improve: A stronger model can outgrow old prompts, restrictions, and permissions, which means an agent can fail not only when software degrades, but also when the underlying model gets better.
Agents inherit the mess of the business around them: Stale wikis, outdated CRM fields, changed dashboard definitions, and obsolete SOPs become dangerous when an agent turns that bad context into polished summaries, recommendations, and actions.
OpenAI and Anthropic are winning by maintaining the workbench, not just the brain: Nate points to Codex and Claude Code as examples where terminal access, browser use, plugins, approvals, sandboxing, logs, and memory are being actively tuned as models and real work evolve.
Every serious agent needs five maintenance checks: He recommends regularly reviewing what the agent reads, what it can touch, what job it is supposed to do, what proof it must return, and whether the output still creates real value after human review.

The Breakdown

Vercel improved a sales agent by deleting 80% of its tools, and Nate B Jones argues that this is the real lesson most agent builders are missing. The hard part is not building an agent, but maintaining the harness around it as models improve, workflows drift, and stale systems turn convincing outputs into business risk.