What is an AI agent (and what it is not)

"AI agent" is the most overloaded term in software right now. Ask three vendors what is an AI agent and you get three answers, usually inside a pitch deck. Here is the working definition we use: software that receives a goal, works out the steps to reach it, takes actions in other systems, and adjusts based on what comes back, without a human directing every move. Everything below follows from that definition, including the cases where you should not build one.

Three things an agent is not

• Not a chatbot. A chatbot answers one prompt at a time inside a conversation window. It does not act. An agent plans and executes across systems, and a conversation may never be involved.
• Not a workflow script. Zapier-style automation runs fixed steps in a fixed order. It breaks the moment an input is shaped differently. An agent has a reasoning step that decides what to do with the unexpected case.
• Not a model. GPT, Claude or Gemini on their own are reasoning engines. An agent is the loop around the model: tools, memory, permissions, retries, and a stopping rule.

If a vendor cannot tell you which of the three they are actually selling, that is your first warning sign.

What an agent actually does

Every agent is some version of the same loop, with three parts:

• Perception: it reads inputs. Documents, API responses, database rows, calendar events, messages.
• Reasoning: a language model interprets the input and picks the next step from a defined set of tools.
• Action: it calls an API, writes a record, sends an email, routes a case, runs code. Then the loop repeats.

An agent processing mortgage applications reads a document, extracts the fields, calls a credit-check API, compares the result against policy rules, routes a decision to a review queue, logs every step, and moves on. Nobody clicks buttons between steps. The value is in the reasoning step: when a document arrives in an unfamiliar format, the agent adapts instead of crashing. That is also where every hard engineering problem lives.

Where agents fail in production

We have seen enough agent builds now to be blunt about the failure modes:

• Long task chains compound errors. If each step is right 95% of the time, a 10-step chain is right about 60% of the time. Narrow scope is not a limitation, it is the design.
• No tools, no agent. An agent with read access to your CRM but no write access cannot update your CRM. The tool layer is most of the build.
• Every reasoning step costs a model call. A high-volume agent that reasons on every record can cost more per month than the people it replaced. Batch, cache, and route the easy cases around the model.
• Accountability does not transfer. Credit decisions, medical routing, payouts: the agent prepares, a person signs. Treat human review as a feature with an owner, not a fallback.

Four tests before you build one

Run the problem through these before anyone writes code. If it fails more than one, you want a workflow tool or a plain integration, not an agent.

• Variability: do the inputs genuinely differ from case to case? If a template handles 95% of them, a template wins.
• Cost of a human: is a person spending hours a day on this? Agents earn their complexity on work that is expensive and repetitive at the same time.
• Measurability: can you state what a correct outcome looks like, per case, so you can score the agent weekly? If not, you cannot tell whether it is working.
• Decomposition: does the task break into 3-6 steps with a known tool at each step? If the steps are fuzzy, the agent will be too.

What we built, and how narrow we kept it

For thedouble.ai we built an AI identity platform. A professional sets up a digital version of themselves that answers questions, captures leads and books meetings around the clock. A new profile goes live in about 10 minutes and then answers 24/7 across channels from one platform. It is an agent by the definition above: it perceives a visitor's question, reasons over the person's own material, and acts by booking or capturing a lead. It is also deliberately narrow. Answers are retrieval-grounded in what the person actually wrote, guardrails keep it on-brand, and anything that needs a real decision is routed back to the human.

For Greenwin, an insurer, the AI layer reads incoming policy and claims documents, fills the records, scores its own confidence, and sends anything uncertain to a person. It is a small loop rather than a free-roaming agent, and that is why manual data entry fell by about 75% and processing got roughly 5x faster without creating a new category of silent errors.

The pattern in both: a tight loop, a short list of tools, and a human on the path for anything that carries real consequences.