Here's where most of the AI conversation goes wrong, and the one rule that keeps an agent worth running. This page is about the leash: the discipline of letting an agent do the work while a human stays on the decisions that carry consequences.
A real agent can act. That's not the same as letting it act on everything. The whole craft is deciding which moves it makes alone and which ones it has to stop and ask about. We call that the leash: a human kept on the calls that matter.
Automate the work. Don't automate the accountability.
In the three agents we build, starting with resume screening, the leash sits in the same place every time: the agent does the tireless 90% (the reading, the scoring, the reformatting, the ranking) and a human owns the moment of consequence. The agent never decides who gets the job. It never sends a candidate to a client without a person saying yes, send it. It clears the path; you choose who walks down it.
We're not apologising for that. It's the design. An agent on a leash is an asset. An agent let off it, making consequential calls, unwatched, at speed, is a GDPR fine and a furious client waiting to happen. Every time the agent in this book hits a moment that actually matters, you'll see the leash. That's deliberate, and you'll get tired of me pointing it out.
Once you know the four parts that make an agent and the leash, you can spot the con.
The market is flooded with products waving the word agent around. Gartner gave the practice a name, agent washing, and a brutal scorecard: of the thousands of vendors claiming to sell agentic AI, only around 130 were judged to be the real thing. The rest are, in the phrase we'll keep coming back to, a chatbot in a trench coat: a talker dressed up as a doer, hoping you won't check whether it has hands.
The tell is simple. Ask what it actually does on its own. Does it touch your systems, or only talk about them? Does it finish a task end to end, or hand you a to-do list and call that automation? Can you see it working, step by step, or is it one confident answer you're meant to swallow on faith? A real agent acts, grounds itself in real facts, loops until the job's done, and stays on a leash. A trench coat answers, then leaves the work to you.
It matters commercially because the failure rate is real money. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and a great many of those will fail not because agents don't work, but because what was bought was never an agent in the first place. You can't operate a costume. There's nothing inside it to keep running.
Four parts: a reasoning engine, hands, real facts, and a working rhythm. One leash, kept on every consequential move. That's an agent: the junior teammate who finishes the job, not the search box that describes it.
And none of it is magic. There's no mystery hiding in any of the four parts, and over the next chapters we'll build every one of them in the open, in real C# against real ATSs. You'll see exactly what's under the hood, which is the only way to know, for sure, whether the thing you're being sold has hands or just sleeves.
Next: the toolkit, the handful of pieces you'll actually need to build one.
Download the full PDF for free?