LLM ToolsA learning lab

Field guide / 01

From chat to working systems

Understand the pieces behind an AI application—and learn how to judge what it can actually do.

4 minute read · Reviewed October 2026

From replies to actions

A large language model (LLM) generates text from the context it receives. An application can also give it tools: searching documents, editing a file, or running a calculation. The surrounding software decides what it can access and how the results are checked.

That distinction helps you evaluate an AI product. Ask what the complete system does, what information it uses, and where a person reviews its work.

Chat assistants
Useful for explanations, drafts, and back-and-forth exploration. Some chat products also include tools, so a chat interface alone does not tell you how much a system can do.
Workflows
Follow a defined sequence, such as extracting information from a document, checking required fields, and preparing a draft. Predictable tasks often fit this approach well.
Agents and harnesses
An agent can choose its next step using a model and tool results. A harness supplies the context, tools, permissions, and execution loop. More autonomy creates more decisions to verify.

These terms overlap across products. Anthropic’s guide to agent systems gives a useful distinction between predefined workflows and agents that decide how to proceed.

Four questions for any AI system

  1. What is the task? Specify a useful result: a summary with source references, a tested code change, or a draft reply.
  2. What is in context? Check which files, messages, retrieved documents, and instructions the model can actually see.
  3. What can it do? Reading a document and publishing a response require different permissions.
  4. What counts as evidence? Compare summaries with originals, open cited sources, and run the checks for a code change.

Connecting more tools does not automatically improve an answer. The Model Context Protocol (MCP) standardizes connections between AI applications and external systems; the application still needs appropriate permissions and review.

Start with a small experiment

The six playground tools use local rules and prepared examples. They do not call a live LLM. Use them to explore the interface and the limits of simple methods, then apply the same questions when evaluating a real model.

Next, explore coding harnesses for project work or OpenClaw for an example of an agent connected to messaging and recurring tasks.

Go deeper