Published: 2026-06-20
Analysis & perspective

Why Better Models Can Break Your Agents: The Case for Harness Maintenance

Chapters / key moments (click to jump — plays here on the page)

Nate B Jones uses Vercel’s sales agent — which got better after the team deleted 80% of its tools — to argue the real 2026 agent story is maintenance, not building. The "harness" (the workbench around the model: sources, memory, tools, permissions, proof, stop conditions) must be continuously pruned and rebuilt, because agents break in two directions: the world around them drifts, and the model inside them improves.

Source video

"Don't build more AI agents until you watch this" by Nate B Jones — Watch on YouTube →

Key Takeaways

Vercel studied a top sales rep’s real workflow, built an agent around it, then improved it by removing tools — "the beginner instinct is to add; the maintenance instinct is to ask what to remove."
Agents break when the model gets better, not just worse: a harness built to constrain a weak model can trap a stronger one, or hand a now-capable model too much reach.
Agents inherit the crud of the systems around them — a stale wiki or a drifted CRM field that’s merely annoying to a human becomes dangerous when a proactive agent acts on it.
Both OpenAI (Codex) and Anthropic (Claude Code) are winning by maintaining the harness — terminal, browser, computer use, memory, approvals, sandboxing, logs — not just shipping a smarter model.
Everyone has a "harness" whether they call it that or not: your project folders, memory, source docs, approval habits and verification loop. Ask what yours is, how you ship it, and what you’ll need to delete later.
A five-point checkup for any serious agent: what is it reading, what can it touch, what is its job now, what proof does it bring back, and is it still delivering value?

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Reproducible steps

Five checks for any serious agent
1) What is it eating: are its sources current and did the workflow move? 2) Test its reach: read only, draft, create tickets, post, update records, spend or publish? 3) Check its job: don't let it change from summarising to planning silently. 4) Check the proof: links to tickets and sources, what it checked and what it couldn't access. 5) Check its value: does anyone read it, does it save time after review, should it be rebuilt or retired?
Prune as well as add
His example is Vercel's inbound-lead agent, which he says improved after the team removed roughly 80% of its tools (his account of Vercel's story; check Vercel's own write-up). The team first studied one of its best reps to build the agent around the real workflow, with human review.
Decide how much harness you own
A light custom harness is instructions, memory, source folders and a short list: the sources, the job, what it can't touch, the proof needed and when a human decides. A deeper one adds data feeds, review screens, permission levels, logs, model choice and escalation, and with it long-term upkeep.

Gotchas

Agents can break when the model improves: a rule or tool that helped a weaker model can trap or confuse a stronger one, and broad access that was safe with a clumsy model may be unsafe with a better one.
Stale wikis, prompts and definitions are inherited by the agent, which keeps producing convincing work from them.