If you've read three AI vendor pitch decks this quarter, you've seen the word "agent" used to mean at least five different things. A chatbot is an agent. A workflow automation is an agent. A retrieval pipeline is an agent. A scripted RPA bot wearing an LLM jacket is an agent. The latest demo of Claude or GPT clicking through a browser is also, somehow, an agent.
When a word does too much work, it stops doing any. We've started keeping a private taxonomy in our Score reports because clients keep nodding along to the word and meaning very different things by it.
Here's the rough division we use, smallest scope to largest.
The conversational helper
This is the thing your CEO probably means when they say "we should add an agent to our website." A bot in a chat window. Sometimes useful, often not, and almost never the most valuable thing you could build. The success metric for these is deflection rate; the failure mode is making your customers angrier with a Turing test before they get to a human.
The single-purpose worker
A bit of code that takes a defined input, runs an LLM call (usually with retrieval), and produces a defined output. "Summarise this email and propose a reply." "Categorise this expense." "Extract the line items from this invoice." These are the workhorses. They look unremarkable. They produce most of the actual value we ship. You could call them functions if you didn't want to make them sound exciting.
The orchestrator
This is where it gets interesting. An orchestrator coordinates several single-purpose workers in sequence, with branching logic, with fallback, with human checkpoints. The agent that triages your inbox, decides which of four downstream workers to hand off to, escalates to a human when confidence drops, and updates your CRM. This is where multi-agent architectures earn their keep. It's also where most of the hard engineering lives.
The autonomous operator
The thing demos love. An LLM clicks around an interface, makes decisions about what to do next, takes actions in the world, runs for hours unsupervised. Anthropic's Computer Use, OpenAI's Operator, the various "browse and book" demos. These are real and getting better quickly. We are extremely cautious about deploying them inside SMBs whose audit logs aren't ready for "an LLM clicked something" as a line item.
The reason this matters isn't pedantry. It's that each of these has a wildly different cost, risk profile, and integration shape. A conversational helper might be a weekend's work. An orchestrator is a quarter. An autonomous operator inside a regulated business is something we won't quote on without a discovery first.
When a vendor says "we'll build you an agent," ask which one. When a board paper proposes "deploying agents across the business," ask which ones, where, and what fails when they're wrong. When a competitor's case study says they "automated with agents," check whether they shipped a single LLM call or actually orchestrated something.
We try to keep the language honest in our own work. The Score report names the type. The Compose phase declares the architecture. The Perform dashboard tells you whether the orchestrator handed off correctly or whether the helper deflected something it shouldn't have.
Words shape what gets built. If everything is an "agent," then anything looks like progress. The corrective is the same one a conductor uses when the brass is drowning out the strings: name the part, listen to it on its own, then put it back in the mix.