The intern-with-keys analogy
A chatbot is the intern who answers your question across the desk: smart, fast, but only words.
An AI agent is the intern you handed the office keys, the company credit card, and a checklist. They can open doors, run errands, file reports, and come back when the work is actually done — not just when an answer is composed.
The leap from chatbot to agent is the leap from talking about the world to changing the world.
The minimum bar for "agent"
A system is an agent when it can:
- Use tools — call functions, hit APIs, run code, query databases.
- Loop — observe the tool's output, decide the next action, repeat.
- Stop — recognise when the goal is met and return.
Take any of those away and you are back at chatbot territory.
The agent loop, explicitly
goal → plan → act (call a tool) → observe (read the result) → decide
↑ │
└────── if not done, loop ─────────────────┘
The model does the plan, act selection, and decide. The runtime does the actual tool calls and feeds results back.
Tools are the leverage
A model with no tools tops out at "things it memorised during training." A model with the right tools can:
- Read live data (database, search, file system).
- Write to the world (send an email, open a PR, ship a deploy).
- Run code (execute, test, iterate on real outputs).
- Call other agents (delegation).
Quality of the tool set predicts quality of the agent more than choice of model.
Where agents fall apart
- Tool sprawl — 50 tools on one agent and the model gets confused about which to use. Group, scope, or split.
- No max-step cap — a stuck agent can spin forever. Always set a hard ceiling.
- Hidden side effects — tools that mutate prod (delete, send, charge) need confirmation gates, not just trust.
- Lossy observation — pasting a 10MB tool output into the next prompt blows the window. Summarise, paginate, or store-and-reference.
Practical autonomy levels
| Level | What the agent does | Where to use it |
|---|---|---|
| 0 | Suggests an action, human runs it | Sensitive prod work |
| 1 | Runs read-only tools alone | Research, summarisation |
| 2 | Runs reversible writes (drafts, branches) | Internal tooling |
| 3 | Full autonomy with audit trail | Background jobs, data pipelines |
Start at 1. Earn the way to 3 with logs and evals.
From the field
The difference between an agent that helps and one that scares clients is entirely in the tools and the guardrails, not the model. My rule on every build: an agent starts read-only, earns write access one tool at a time, and anything irreversible — send, delete, charge, deploy — sits behind an explicit confirmation gate no matter how much I trust the model that day. The other thing demos hide is observation cost: a tool that dumps a megabyte of output into the next prompt will blow your context and your bill. Summarise tool results before they re-enter the loop. Those two habits prevent most agent horror stories.
Custom Claude Code AI Agents & Workflows
Stop doing the repetitive, multi-step work that eats your team's day. This service delivers a working AI agent system that handles tasks like lead processing, data enrichment, content pipelines, and r...
