An AI agent is a chatbot that can act. It uses an LLM as its brain, has access to tools and context, and works in a loop until a task is done – reviewing a briefing, setting its status and drafting the follow-up query to the client, for example.
What is an AI agent?
Where a chatbot only outputs text, an agent can do real work. From the LLM it gets not just an answer but a plan and carries it out step by step. This is called the agent loop: the model reasons, picks a tool, runs it, looks at the result and decides what comes next. The loop runs until the goal is reached. The actual execution happens in a protected environment, the so-called sandbox.
An agent is made up of several building blocks that work together:
- a trigger that starts it (a prompt, a new task, a schedule)
- a system prompt with role, guidelines and goal
- an LLM as its brain
- context, meaning the knowledge it works with
- tools and MCP to act and connect to other systems
- skills to handle tasks consistently
- memory to remember things over time
- a human in the loop who signs off at the important points
Why this matters for agencies
The leap from chatbot to agent is the leap from "AI that spits out text" to "AI that does real groundwork inside the client project".
Instead of everyone prompting on their own, a specialised agent takes on recurring tasks – on the real agency and client context, and with clear permissions.
Typical agents agencies are already talking about:
- Briefing agent: checks incoming briefings for gaps and contradictions.
- Brand-safety agent: compares assets against the client's brand and tone-of-voice guides.
- Costing agent: produces a first quote based on the rate cards.
- Research agent: researches competitors and the market for the pitch.
- Reporting agent: builds the project report automatically every Friday.
An example from agency life
A new briefing lands in the project's inbox. That's the trigger. The briefing agent takes over automatically, sets the status to "in review", reads the document and the stored client context, compares it against your briefing template and finishes by creating a review document plus a draft of the follow-up email. Before that email goes out, the account lead signs it off.
How this connects to other terms
An agent builds on the LLM but goes far beyond the chatbot. To act, it needs tools and MCP; to do so consistently, it needs skills; and to keep you in control, it needs a human in the loop. The whole picture is in the article "From LLM to agent".
In awork, agents run directly on your existing project structures: with automatic access to agency, client and project context, shareable across the team and using the permission system you already have. That's what makes rolling out AI in the agency manageable in the first place.
