85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Glossary

What is an AI agent?

An AI agent is a system in which a language model directs its own steps and tool calls toward a goal, reading each result and choosing the next action until the task is done.

The term is stretched to cover almost anything with a model inside it, from a help widget to a scheduled script. The labs that build the models use it more narrowly. Anthropic's "Building effective agents" (December 2024) separates workflows, where code fixes the path, from agents, where the model decides its own process and tool use. OpenAI's "A practical guide to building agents" (2025) says an agent controls the execution of a task and picks tools within guardrails, and states that chatbots and single-turn classifiers are not agents. Google's agents whitepaper describes a model plus tools plus an orchestration layer running a reason, act and observe cycle.

For a product team the useful test is simple: who decides what happens next? If code decides, the system is a workflow with model steps. If the model decides, reading live results from tools as it goes, it is an agent, and it needs the controls an agent needs.

What an AI agent is made of

  • A model

    The language model that reads the task and the context and decides the next action. OpenAI's guide advises prototyping on the most capable model, setting an eval baseline, then trying smaller ones.

  • Tools

    Functions the agent can call to read data or change something: search a knowledge base, look up an order, issue a refund, run code. Without tools there is nothing to act on.

  • Instructions

    The system prompt and task description that set the goal, the rules and what done looks like.

  • A harness

    The surrounding code that runs the loop, executes tool calls, manages context, enforces permissions and keeps state. Two agents on the same model can behave very differently because of it.

  • Checks

    Evals, traces and approval gates that show whether the agent did the right thing, and stop it before an irreversible action goes wrong.

Where agents sit on the product ladder

Motiontic's thesis describes four levels of AI in a product. L1 is a bolt-on chatbot, the level users learn to ignore. L2 is model calls behind existing features, such as a summarise button, where engagement tends to stay flat. Neither is an agent: a person still decides every next step.

L3 is AI-native flows, such as onboarding that configures itself. These are often agentic workflows, with the path mostly fixed and a model making bounded choices. L4 is the agentic product, where the user states an outcome and supervises while an agent plans, acts and checks. L4 is where AI agents in the strict sense live, and where most of the engineering work is tools, permissions and evals rather than prompts.

When an agent is the wrong choice

All three labs give the same advice: start with the simplest design that works. Google's Cloud Architecture Center notes that predictable tasks, or anything one model call can do such as summarising, translating or classifying, are often cheaper without an agent. OpenAI's guide reserves agents for work where rules engines fail: decisions that need judgement, rule sets too complex to maintain, or heavy unstructured input.

An agent costs more per task than a fixed path, takes longer, and varies more from run to run. It earns that cost when the number of steps cannot be known in advance.

Where AI agents fail in production

  • Wrong tool, wrong arguments

    The agent calls a tool that does not fit or passes an ID it made up. Fewer, clearer tools and argument validation reduce this.

  • Stopping early or never stopping

    It declares the job done before it is, or loops on a task it cannot finish. Explicit done criteria and iteration caps guard against both.

  • Untrusted input steering it

    A document, email or web page it reads contains instructions. Prompt injection is reduced by design, not solved, so what the agent can reach has to be limited.

  • Cost drift

    Every turn adds tokens. Without per-task budgets and traces, spend grows unnoticed.

How Motiontic's engineers apply it

Motiontic's engineers map a process step by step before choosing a design, then give the model control only of the steps that need judgement. Each agent ships with a tool list scoped to the job, an eval set built from the client's real cases, traces for every run, and a human gate on anything that cannot be undone.

What an AI agent needs in production

What an AI agent needs in production
PartWhat it doesWhat breaks without it
Scoped toolsGives the agent the actions the job needs and no moreThe agent reaches systems it should not, or has nothing useful to do
Done criteriaDefines finished in a form the agent can checkEarly stops and confident half-finished work
Iteration and spend capsStops a stuck runRunaway loops and surprise bills
TracesRecords what the agent saw, called and decidedFailures that cannot be explained or reproduced
Evals on real casesScores outcomes before and after every changeRegressions found by customers
Human gatesHolds irreversible actions for approvalRefunds, deletions or emails that cannot be taken back

Common questions

What is an AI agent in simple terms?

An AI agent is software where a language model decides what to do next, calls tools to read or change real systems, looks at the result, and repeats until the task is finished or a limit is hit.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions in a conversation. An agent takes a goal, calls tools to act on live systems and checks its work. OpenAI's guide to building agents states plainly that chatbots are not agents.

Is a workflow that calls an LLM an AI agent?

Not in the labs' sense. Anthropic calls that a workflow, because code fixes the order of steps. It belongs to agentic AI, the wider category, but the model is not choosing its own path.

Do AI agents need a human in the loop?

For irreversible or high-risk actions, yes. For routine low-risk steps, approving every action leads to approval fatigue, so production agents usually approve by risk tier and monitor the rest through traces.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.