Issue 022: Agents in Production

Here is the claim. In 2026 a non-developer can put an agent into a live business process without engineering support, and when that agent fails it will almost always be because the process was never properly defined.

Start with integration, because that was the honest blocker for years. Connecting a model to email, calendars, a customer relationship management (CRM) system, a ticketing queue and an internal document store used to need a developer and a maintenance budget. Zapier alone now connects to more than 9,000 applications, and this week's MCP revision turned the wider connector layer into ordinary infrastructure that runs on standard cloud plumbing. The integration gap has closed for the tools most businesses actually run on.

Second, supervision arrived as a feature rather than a custom build. A year ago, pausing an agent for human approval before it sent an email or moved money meant writing middleware. Approval steps, activity logs and run histories are now shipped in the platforms themselves. The week's regulation is pushing in the same direction: the AI Kill Switch Act would make suspension capability a legal requirement, and the European Union's serious-incident duty makes the audit trail something you have to be able to produce.

Most organisations can name the recurring tasks they want handled. Invoice chasing. Lead qualification. Meeting preparation. Weekly reporting. Very few can produce a written description of how that task is actually done: what triggers it, what the acceptable range of judgement is, which exceptions must go to a person, what evidence gets kept, who owns it when it goes wrong. That document does not exist in most businesses because a human has always held it in their head. An agent cannot read a head.

The strongest counter-argument is that this is all irrelevant in high-stakes work. In financial services, healthcare and legal practice, the tolerance for failure is so low that no amount of process design makes current agents safe to run unsupervised. That is correct, and this week's AISI findings sharpen it: if models will quietly break the rules of a task to complete it, and will not reliably tell you afterwards, then evidence of good behaviour is harder to obtain than the marketing suggests.

But that argument narrows the claim rather than defeating it. In regulated work the answer is a smaller scope and human completion rather than human review, which is itself a process decision. And it leaves untouched the large operational middle where most professional work sits, where an error costs an apology and a correction rather than a fine.

So here is the recommendation. Take one recurring process this week and write it out end to end, in steps, before you open any tool. Mark every point where a person must approve. Name the failure that would embarrass you in front of a customer, and write the rollback for it. If you cannot describe the process in plain language on one page, you are not ready to automate it, and no model release in the next twelve months will change that.

An agent layer sitting on top of the automation platform many businesses already use, sold as Zapier Agents and previously known as Zapier Central. You describe what the agent should do in plain English, connect the applications it is allowed to touch, and it takes action across email, calendars, customer relationship management (CRM) records, documents and more than 9,000 other connected applications. No code at any point.

This issue argues that agents are now a process problem rather than a technology problem. Zapier Agents is the cleanest test of that claim, because nothing in the build stops a non-developer. If your first agent fails here, the reason sits in your process description.

Zapier's own documentation is candid that identical inputs will not always produce identical outputs, so anything irreversible needs an approval step in front of it. Run history and activity monitoring are good, but organisation-wide audit logs sit at the enterprise tier, approval requests are still in beta, and undoing an action the agent has taken in another system remains your job.

Pricing as at 29 July 2026, billed yearly. Paid monthly, Professional is £22.85 ($29.99) and Team is £78.85 ($103.50). Zapier notes the sterling price can move with the exchange rate at the time of payment, and all prices exclude tax. Agents is included on every plan, with the free tier covering 400 agent activities a month.

Try Zapier →

The story of this week is not that agents got better. It is that the excuses ran out. Integration is solved for the tools most businesses run on. Approval and audit are shipping as standard features. Regulators are turning containment into a legal duty rather than a good practice.

What is left is the part nobody enjoys: writing down how the work is actually done, and deciding where a person still has to say yes. That is a management problem wearing a technology badge. It is also good news for anyone who has ever run a proper process mapping exercise, because that skill is about to become commercial again.

Two predictions, both checkable. By 31 December 2026, "agent-ready process" will be the more common phrase in enterprise AI programmes than "AI-ready data". By 31 July 2027, a publicly disclosed agent failure inside a regulated UK firm will trigger formal regulator guidance on the operational governance of autonomous agents.