Agents are sailboats, not apps — the real skill is maintenance, not launch.
Vercel made its sales agent better by deleting 80% of its tools. The lesson isn't that AI replaced a process — it's that useful agents desperately need good maintenance, and pruning is the upgrade.
The beginner adds. The maintainer asks what to remove.
Vercel studied one of its best sales reps — not the paper workflow, the observed one. What did the rep ignore? What made a lead real? What research happened before a reply? When was a "sales" question actually support? Where did a human still need judgment? They built an agent around that real behaviour: it filtered inbound messages, qualified leads, researched companies, drafted responses, routed support away from sales. A human reviewed the work.
That alone is a clean story. The important part came after the agent existed. It didn't improve when the team piled on tools. It improved when they took tools away — and not for context-window reasons, the usual excuse. Pruning was the upgrade.
Useful agents desperately need good maintenance. That is the real lesson — not "AI replaced the sales process."
Agents break in a way ordinary software never has
We're used to software breaking when it gets worse. Agents also break when the model gets better — a strange new maintenance problem. Four principles frame the discipline.
The model underneath is moving
It keeps getting better at tool use, reasoning, messy instructions, remembering what matters. Yesterday's harness can become wrong overnight — a rule that protected a clumsy model can trap a better one.
Agents inherit all the crud
Wikis go stale. CRM fields drift. A dashboard still says "activation" after the team redefined it. An agent doesn't sit on stale truth — it produces convincing work from it.
The biggest labs already know
Codex isn't strong only because the model is strong. OpenAI keeps maintaining the surface: terminal, browser, plugins, memory, approvals, sandboxing, logs. Anthropic makes the same bet with Claude Code and Cowork.
So ask: what is my harness?
Folders, memory, prompts, source docs, approval habits, file rules, your verification loop, your demand for proof. The mature question isn't "can I build an agent" — it's whether you can keep the setup healthy.
Agents are less like apps and more like sailboats
People call the setup around an agent a harness. If that feels too technical, call it a workbench. The agent is the worker; the harness is what it reads, what it remembers, which tools it can touch, what it's allowed to change, what proof it must bring back, and what stops it when the work gets risky.
A sailboat isn't maintained because it was badly designed. It's maintained because it lives in motion. The weather changes, the lines loosen, salt gets into everything. Agents live in motion too.
The world drifts
Docs go stale, processes move, sources become misleading — and the agent keeps working anyway.
The model improves
New capability turns a once-safe permission into too much reach, or a once-needed guardrail into a cage.
Maintenance is the work of keeping the harness fit between those two moving things — the world drifting and the model improving. That's closer to vehicle maintenance than to anything software has shown us in a long time.
Better agents build better harnesses build better agents
There are really two teams building world-class harnesses right now: Anthropic and OpenAI. If the model can help you ship, test, refactor, observe, and prune the harness, capability gain compounds. The winners aren't the ones who build the perfect wrapper once — they're the ones who keep rebuilding the wrapper as the model and the work change.
A better model helps build a more effective harness
The model itself can draft, test, and refactor the surface around it — not just the product logic.
A better harness makes the agent more useful
Cleaner tools, sharper memory, tighter permissions — the agent produces better work from a better workbench.
A more useful agent touches more real work
More real work creates pressure to improve the harness again — the flywheel ignites.
More real work creates pressure to improve the harness again
The loop is not a circle you walk once. It accelerates. Each pass raises the bar for everyone else.
Building your own agent now means choosing not just a model, but how much harness maintenance you own versus outsource. A light harness can be a clean set of instructions, memory, and source folders. A deep one adds data feeds, review screens, permission levels, logs, escalation paths, approval rules, and a plan for when the model changes. The more custom the harness, the more you own the upkeep.
Five things to check on any serious agent
None of these started with AI — but a machine that produces convincing work from a mess makes every one of them urgent.
What is it eating?
What is it reading, and are the sources current? Did the workflow move? Did a new source become important — or an old one become misleading? Stale input is the quietest failure.
What can it touch?
Read only? Draft? Create tickets, post to Slack, update records, spend money, publish? A permission harmless for a weaker model may be far too broad for a strong one — and a restriction built for an unreliable model may be holding a better one back.
What is its job?
Still a summary agent — or quietly becoming a planning agent because it now can? Don't let the job change silently. If the role should grow, change it on purpose.
Where is the proof?
"Customers are frustrated with onboarding" isn't enough. It should link the tickets, quote the customer language, cite the sales notes, and say which sources it checked and which it couldn't reach — a linkable trail a human can inspect.
What is its value?
Does anyone read the output? Does it change the work or save time after review — or create a new pile? Has the model improved enough that the agent should be rebuilt, or the business changed enough that it should be retired?
An agent doesn't have to fail loudly to become dangerous. All it has to do is keep working on a stale truth — and it will start to haunt the business.
The Maintenance of Everything
Stuart Brand, on sailboats, vehicles, corrosion, and the work that keeps important systems alive after the launch moment is over. Simplicity, he argues, is a key to maintenance — which is exactly the Vercel lesson.
If you take one thing from this guide: agents are not launched, they are maintained. The teams that win are the ones who keep the workbench clean, the tools pruned, and the harness fit — every week, not once a year.