Start with the work, not the agent

The most impressive AI projects we have seen delivered the least. They start with the agent. Someone builds something clever, demonstrates it, and then goes looking for a place to put it. They launch well and quietly fade, because they were never anchored to anything a team actually needed. The projects that last do the opposite, and they often land somewhere people do not expect: you did not need an agent at all.

A task with fixed rules and tidy data needs plain automation, not an agent. Pulling a few fields out of a document needs a simple tool, not an agent. An agent earns its place only where the work has several steps, real judgement, and enough variation that a fixed process cannot cope. Everything else is over-engineering, and over-engineering is how budgets get spent on things nobody uses.

So the first question is not what can this agent do. It is how does this team work, and where does that work break down. Take an estate agent. A sale is agreed, and then the handoff to conveyancing stalls: the same details keyed again into another system, chased by email, left sitting in someone’s inbox over a weekend. Map the day like that, step by step, and the opportunities are obvious before any technology is chosen. The stalls, the re-keying, the skilled people spending hours on work that does not need them.

You are not asking how to make an agent impressive. You are asking how to make a person faster and freed from the parts of the job that add nothing. That is a small change in what you ask first, and it changes what you end up with. It is the difference between AI that gets adopted and AI that gets abandoned.

Once an agent has earned its place, building it is mostly assembly. You write down how the job is done, hand over only the tools that job needs, and that pair is the agent. One agent does one substantial piece of work. Chain those together and you have a run of work that used to take a person all morning.

None of it has to be built from scratch either. Every part above is something Claude already gives you.

The part worth real thought is risk, and risk here is not about how clever the agent is. It is two plain things: what its tools can reach, and how much it does with nobody looking. Those two set how much of a person you keep in the loop, and how closely you watch it.

The deeper payoff comes after you build. Treat an agent like a new member of the team: give it clear expectations, watch it, and improve it on feedback. Set it up from the start so every step it takes leaves a record of what it did and how well it did it.

That record is not just for improving the agent. It shows you how to improve the work. The points where an agent struggles are almost always the points where the process underneath is unclear or inconsistent. So the agent becomes a way of understanding your own business better, and each round of refinement improves both the tool and the process it sits in. That advantage builds on itself, and you only get it if you watch and adjust rather than setting the thing live and walking away.

The figures above are the machine side of this: skills, tools, agents and the controls around them. Anthropic gives the human side a name, four habits that run alongside the machinery, and they line up with the order this piece has been arguing for.

Start with the work, not the technology. Build only where an agent genuinely helps, and say where it does not. Then treat going live as the start of the loop, not the end of the job. Done this way, AI stops being a set of clever tools looking for a purpose and becomes a way of making a team better at what it already does.