We first wrote this post in December 2023, when “AI agent” was a term you had to explain. Three years on it needs no introduction, but it does need rescuing. Every product now calls itself an agent: chatbots, autocomplete, a scheduled script with an API key. The word has been stretched until it covers almost anything, which means it tells you almost nothing.
So here is the definition we actually use, revised after three years of building and running agents in production.
A definition that does some work
An agent is an autonomous system with decision-making capability. The old word for this is agency: the system does not just execute a fixed sequence, it decides what to do next based on what it finds.
That short definition rules out most of what gets sold under the label. A chatbot is not an agent; it waits for you. A prompt with a nice interface is not an agent; it has no decisions to make. And MCP, which now appears in every AI product announcement, is not an agent either. It is plumbing: a standard way for a model to reach data and tools. Agents use MCP. The protocol no more makes an agent than a telephone makes a negotiator.
Agency is a spectrum, not a switch
The more useful question about any system is not “is this an agent?” but “how much agency does it have?”
At one end sit hardwired pipelines: a fixed chain of API calls with a language model doing one task at one step. Reliable, cheap, and not remotely autonomous. At the other end sit systems that plan: given a goal, they decide which steps to take, which tools to use, and when the job is done.
Most of what we built in our first year lived at the hardwired end. Most of what we build now lives at the planning end, because the models have improved enough to be trusted with decisions, not just tasks. But the spectrum is the point: placing a system honestly on it tells you what it will cost, how it can fail, and how much supervision it needs. Vendors who call everything an agent are hoping you will not ask.
What a production agent actually looks like
Strip away the branding and our agents share one anatomy: a layer of judgement and a layer of access.
The access layer is the easy part to describe: connections, usually via MCP, to the systems that hold the client’s material – archives, databases, monitoring feeds, document stores.
The judgement layer is what makes the output usable. It encodes who the audience is, what language the organisation writes in, what must never be invented, and what a good result looks like. Data access without judgement produces plausible rubbish; judgement without data access produces an eloquent intern with no files.
Two habits from production are worth passing on.
Agents should check their own work. Our translation agent runs two passes: one to translate, a second to verify the translation against the source. The second pass catches the errors the first one cannot see, for the price of one more model call. We now treat a verification step as the default in any agent whose output goes to a reader.
Long chains break; parallel workers do not. An early version of one reporting agent ran as a single sequential chain: gather candidates, investigate each in turn, write up. One slow connection mid-run pushed the whole job past its time limit and killed it, taking fifteen minutes of work with it. The rebuilt version uses delegation: a first agent selects what deserves attention, then independent sub-agents investigate each item in parallel. It is faster, and when one worker fails it fails alone, cheaply, and can be retried. Fragility, it turns out, is an architectural choice.
Deciding by evidence, not by blueprint
The honest answer to most architectural questions is that we did not know in advance. We treat each design as a conjecture and let production evidence settle it: the sequential chain was a reasonable guess, the failure was the refutation, and delegation was the revision. A design we cannot imagine failing is a design we have not thought about hard enough.
The same discipline applies to the first question of any engagement: whether an agent is the right tool at all. Not every problem needs one. If a task is fixed, frequent, and identical every time, a plain automation is cheaper and more dependable, and we will say so. Agents earn their keep where the work requires judgement on every run: what matters in today’s coverage, which of these documents answers the question, what this draft should say to this audience.
That is what we mean by AI agents: systems with enough agency to be delegated real work, built with enough humility to prove they deserve it.
Originally published 13th December 2023; rewritten July 2026 to reflect what three years of production deployments have taught us.
Leave a Reply