the fence is the job

You might think the skill worth money in the agent era is building more capable agents.

Fair enough.

That's what every demo wants you to believe.

But then why does the biggest domain on Anthropic's new exam focus on handoffs, escalations, and what agents should not do on their own?

I'll tell you about two documents that, together, tell the real story.

One is Anthropic's new certification for people who architect agent systems instead of just prompting them.

It's called CCA-F, Claude Certified Architect Foundation, pitched like a 301-level exam for people who've spent real time in production.

I sat it. I passed it. The certificate is mine, so none of this is secondhand.

This isn't a certification flex. I'm telling you what showed up on it, because the pattern matters more than the credential.

Here's what's on the exam:

Five domains total. The biggest one, 27% of the material, is agentic architecture and orchestration.

Not model choice. Not prompt design.

And the questions are scenario-based: here's an architecture, what's likely to go wrong?

Production teams are already pricing this exact judgment call into retainers, whether you sit an exam for it or not.

There's a full section on reliability and escalation patterns, and it's very direct about one thing: you cannot trust an agent's self-reported confidence score.

Ask it how sure it is, and that number may have nothing to do with whether it's right.

The other document is a startup playbook for selling agents as labor, not software.

Different audiences. Same message.

And the overlap is the whole point.

The startup playbook is working the same problem from the other side.

It lays out a four-rung ladder for how much autonomy to hand an agent: draft and approve, triage, coordinate, and bounded action, the narrow stuff you can safely trust it with.

A refund under $50, clear rules, small blast radius.

And the advice to founders is blunt: earn the autonomy. Don't give it away just because the demo looked good.

You could argue capability is still the main lever.

Every new model release expands what agents can do, so why wouldn't it be?

But capability is also the part that keeps getting cheaper.

What doesn't come free is knowing where the agent should stop.

Both documents above kept circling back to that exact question: one prices it in exam points, the other prices it in a retainer, something like $1,000 a month for a workflow a customer used to run by hand.

So, here's the fence.

That four-rung ladder is really just a fancy fence-drawing exercise. Four questions, asked of any system before you let it run unsupervised.

What wakes it up?

What can it do without you watching?

Where does approval happen?

And when does a human have to step in?

Answer those honestly for your own setup, and you've basically sat the reliability section of the exam without opening the textbook.

Everything else is syntax.

I keep thinking about a restaurant host.

Anyone can learn to answer "what time do you open?"

The real job is knowing when the kitchen stops taking orders even if the sign says otherwise, which tables can fit a stroller, and when a call goes to the manager instead of being handled at the host stand.

Give an agent that phone, and it should earn each of those decisions one rung at a time.

Never all at once just because the demo went well.

And here's the good news:

You don't need Anthropic's blessing or a signed contract to start drawing your own fence.

I have the paper now, and imho it mostly confirms what production already teaches: every system you're running today, a support bot, a code review agent, an internal tool, is already sitting inside a harness.

Whether you've written the rules down or not.

Written down or not just means the fence exists whether you drew it on purpose or backed into it after something broke.

One of those is a lot cheaper than the other.

Buried in the same material is Anthropic's own advice that most agent problems should start as a plain workflow, not an agent.

Which is just another way of drawing a fence.

Don't give away autonomy the task never asked for.

So no, the valuable skill was never "can you make the agent do the thing?"

It's whether you know where the thing should stop.

Anthropic prices that in exam points. Customers price it in retainers.

Either way: the fence is the job.

Draw it, defend it, and people trust you with the rest.

So before you ship the next one: pick one system already running in prod and answer the four questions for it.

Not hypothetically. If you can't say where approval happens, that's not a philosophy gap. That's the fence you haven't drawn yet.

If you're building agents right now, bookmark this before your next autonomy call.