What is an AI Agent? An LLM that can take action – not just answer questions

The short answer

An AI agent is an LLM wired up to tools so it can take action in the real world, not just produce text. A regular chatbot reads your message and writes a reply. An agent reads your message, decides which tool to call (a database lookup, a calendar booking, an email draft, a workflow trigger), calls it, looks at the result, and then either takes another action or replies. The two clearest examples we have shipped: the innatura email agent (classifies inbound enquiries and drafts replies for human review before any response is sent) and the Mother Earth AI voice agent (a real-time spoken interface plugged into backend tools).

What an AI agent actually does, in three steps

1. Read – The agent receives an input: an email, a spoken question, a chat message, a triggered event.

2. Decide and act – The LLM, with a list of available tools, decides which tool to call. Tools are functions: "look up this customer", "draft this reply", "create a calendar event", "post to this Slack channel", "trigger this n8n workflow". The agent calls the tool, reads the result, and may chain another action.

3. Respond or hand off – The agent either replies directly to the user or hands the result to a human for review. For anything that goes to a customer, member or beneficiary, we build mandatory human review into the workflow.

When to use an agent (and when not to)

Use an agent when the task involves taking action, not just looking up information. Drafting replies, routing inbound enquiries, booking calls, triggering workflows, querying multiple systems and combining the results – all agent territory. innatura and Mother Earth AI are the clearest worked examples.

Do not use an agent when a simpler tool fits. If the answer is "look up this fact in our knowledge base", a RAG chatbot is faster, cheaper and more predictable. If the action is "always do X when Y happens", a deterministic n8n workflow without an LLM in the decision path is more reliable.

Always design against the failure mode. An agent that takes action autonomously is an agent that can do the wrong thing autonomously. Mandatory human review on outbound communication, audit trails as a default, and grounded data instead of free-text guessing.

What agents look like in production

innatura (email agent) – inbound donation enquiries arrive in a shared inbox. The n8n workflow reads each email, classifies it by type, retrieves the relevant context, and drafts a reply. A human reviews and sends. The agent handles the reading and drafting; no reply leaves without a person in the loop. Mother Earth AI (voice agent) – a real-time spoken interface for an AI persona based on the Allgemeine Erklärung der Rechte von Mutter Erde. The agent handles speech recognition, tool calls and response generation in a single loop. It demonstrates what an AI agent looks like when the "tool" is an external API and the input is audio rather than text. Both use the same underlying pattern: LLM + tools + mandatory human checkpoint for anything consequential.

Why N3XTCODER

We bring a decade of impact-tech experience and over 160 AI projects since 2019. Through our free AI for Impact course, more than 100,000 people have learned to use AI for the common good.

• innatura – AI email agent classifying donation enquiries and drafting replies with mandatory human review. The production example of an agent with a strict human-in-the-loop step.
• Mother Earth AI – real-time spoken interface using a fine-tuned model for voice and a retrieval layer for facts. The production example of a voice agent wired to backend tools.
• Default stack: n8n in Berlin, Qdrant in the EU, Azure OpenAI via Microsoft EU Sovereignty.

Honest constraints

Agents are only as reliable as their tools. If the database query returns stale data, the agent will draft a reply using stale data. Garbage in, garbage out applies at every step in the chain.

Autonomous action without oversight is a liability, not a feature. We build mandatory human review into every outbound step. An agent that can send without a human check is an agent that will send the wrong thing.

Multi-step chains amplify errors. Each tool call is a potential failure point. Error handling, dead-letter queues and alerting are not optional – a chain that silently fails partway through is worse than one that fails loudly at step one.

Frequently asked questions