Autonomy is a system’s ability to choose and carry out steps toward a goal without asking the user to decide every step. An assistant typically handles a request and stops; an agent can use tools, observe the results, and decide what to do next. The distinction is a spectrum, not a label determined by whether the system ever asks permission. section 1.1.1 section 3.2.4
Table 1.1 The four LLM interaction patterns covered in this book
| Pattern |
Approval model |
Autonomy level |
Typical use case |
Example platform |
|---|---|---|---|---|
| Direct LLM chat |
No tools invoked; the model only generates text |
None |
Q&A, drafting, brainstorming, learning |
Original ChatGPT (2022), early Claude (2023) |
| Tool-augmented LLM |
Implicit per call; a user prompt triggers a single tool |
Low |
Image generation, web search, simple lookups |
ChatGPT with DALL·E, Gemini with Google Search |
| Assistant |
Per-task approval; the user confirms each action |
Medium |
Pair programming, cited research, document editing |
GitHub Copilot Chat, Cursor, Claude in Excel |
| Agent |
Goal-level approval with gated high-stakes actions; the user sets the objective, and the agent decides the steps |
High |
Multistep research, repo-wide refactors, browser automation |
Claude Code, OpenAI Operator, Devin |
Compare direct chat, tool-augmented LLMs, assistants, and agents by autonomy level and approval model. The progression shows how a user can move from requesting an answer to setting a goal whose steps the system selects. section 1.1.1
Research as a multistep goal
A user can ask for research rather than prescribe every lookup. An agent may select a tool, inspect its result, and choose another step. That freedom also compounds mistakes: one incorrect tool call can lead to several more before a person notices. section 1.1.1 section 3.2.4
Grant autonomy within explicit limits:
- Let low-risk steps proceed, but require confirmation for high-stakes actions such as sending email, making purchases, or deleting records. Permission prompts alone do not make a system an assistant; agents can also pause for approval. section 1.1.1 section 3.2.4
- Restrict available tools, isolate filesystem or shell operations, validate tool results, and cap rates or budgets so an erroneous loop cannot run indefinitely. section 3.2.4
- In multi-agent workflows, start with the simplest coordination pattern that works. Balance the agent’s freedom to decide against the system’s ownership of what it is allowed to do. section 4.1.3
