AI Agent

How it works

1. The agent gets a goal ("what important happened in the lab today?") and a list of tools describing what each one does.
2. The model decides which tool to use and calls it, e.g., a database query.
3. The tool's result goes back to the model. The model assesses whether it is enough or if another step is needed.
4. The loop continues until the model considers the task done or the step limit is reached.

Agent or automation?

• Automation (workflow): steps are predefined by a human. The model can perform a single step, like writing a summary, but does not decide what happens next. Predictable, cheap, easy to check.
• Agent: steps are chosen by the model. Flexible, but more expensive, slower, and harder to control.
• Rule of thumb: if you can draw a flowchart of steps, build an automation. Use an agent where subsequent steps cannot be predicted.

How our lab does it

• The lab runs 13 processes in n8n and almost all are automations, not agents: source collection, evaluation, writing, publishing, and alerts have fixed steps.
• The lab CEO is an agent, whom we chat with on Telegram. It has 5 tools: lab pulse, lab status, backlog, Google Calendar, and SQL database queries. It decides on its own which ones to use to answer a question.
• The agent does not publish on its own. Every publication waits for a human click. The agent can prepare, but decisions with consequences belong to humans.
• Lesson from the firing range: process "success" is not the same as a job well done. We check the outcome (whether the report arrived at 9:00 AM) rather than just the status.

What our database shows

Agents are currently one of the top three AI research topics: in week 43, they accounted for 26.3% of papers in the cs.AI category on arXiv. In the lab database, the word "agent" appears in the title of over 2.7k texts. A common theme: agents are more easily manipulated than a regular chat (sycophancy, indirect prompt injection) because they do more on their own.