Act: Taking the Action
Thinking without acting accomplishes nothing. An agent that perfectly understands a situation but never does anything about it is useless. The act stage is where the loop produces real-world effects — where a message gets sent, a search gets run, a file gets written, or a robot arm moves. This is the moment the agent's reasoning meets reality.
What Counts as an Action?
An action is any operation the agent performs that changes something — either in the world or in its own context. Actions fall into several broad categories. Output actions produce a result visible to the user or another system: writing a response, generating an image, producing a report, or speaking a sentence. Tool-call actions invoke a specific capability: running a web search, querying a database, reading or writing a file, sending an HTTP request, or executing a line of code. Environmental actions affect physical reality: a robot picking up an object, a drone changing course, a smart thermostat adjusting temperature. Internal actions change the agent's own state or context: adding a note to memory, setting a flag, deciding to break a task into sub-tasks.
In the agent loop, an action is any operation the agent performs that changes something — in the environment, in a tool's output, in a conversation, or in the agent's own internal state.
Tools: Extending What an Agent Can Do
A language model on its own can only produce text. But text is often not enough — sometimes the agent needs to run a calculation, look up a live fact, or write to a file. This is where tools come in. A tool is a capability that has been explicitly connected to the agent and made available for it to invoke. Common tools include web search, code execution, file reading and writing, API calls to external services, and memory retrieval. When an agent calls a tool, it provides a structured input — called the tool arguments — and the tool returns a result that flows back into the agent's context as an observation. Designers decide which tools to give an agent, and this decision directly shapes what the agent can and cannot do. An agent with no tools can only generate text from its training. An agent with the right tools can interact with virtually any digital system.
Tools are pre-built capabilities — web search, code execution, database queries, API calls — that designers attach to an agent. Calling a tool is one of the most common types of action in modern AI agent systems.
Complete the sentences about agent actions and tools.
Actions Have Consequences — and That Demands Care
Unlike a human who pauses to double-check before clicking 'Send All,' an agent can execute actions at machine speed. This efficiency is also a risk. If the reasoning stage produced a flawed plan, the action stage will carry it out faithfully — and some actions cannot be undone. Sending an email cannot be recalled. Deleting a file without a backup is permanent. Placing a large financial order is binding. This is why thoughtful agent design includes safeguards: confirmation steps before irreversible actions, sandboxed environments where actions are simulated before being executed for real, and human approval gates on high-stakes decisions. The act stage is powerful precisely because it reaches into the real world — and that power demands proportional responsibility.
Some actions — deleting data, sending messages, making purchases — cannot be undone. Well-designed agent systems include confirmation checks or sandboxed testing before executing irreversible actions.
Match each action type to a concrete example of that action.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
What is a 'tool' in the context of an AI agent?
Why do well-designed agents include confirmation steps before irreversible actions?
Design an Agent's Toolkit
- Step 1: You are designing a 'homework research agent' that helps students find and organize information for school projects. List five tools you would give it and explain what each tool does.
- Step 2: For each tool, write one example of the arguments the agent would provide when calling it — for example, a web search tool would receive a search query string as its argument.
- Step 3: Identify which of your five tools could cause irreversible effects if used incorrectly. What safeguard would you add for each risky tool?
- Step 4: Are there any tools you deliberately chose NOT to give this agent? List two tools you excluded and explain why those capabilities would be inappropriate or dangerous for a homework agent.