Every few months a new word arrives for the same idea. "Copilot", "assistant", "agent". Underneath, the agents that actually work in production share a surprisingly small shape: a loop around a model that can call tools.
The core loop
The model never "does" anything. It proposes. Your runtime decides whether to execute, feeds the result back, and asks again.
That's it. Everything else — memory, guardrails, multi-agent setups — is decoration on this loop.
The four parts that matter
| Part | What it does | Where it goes wrong |
|---|---|---|
| Model | Reasons, picks the next step | Overconfident on missing context |
| Tools | Touch the real world (APIs, files, DBs) | Vague descriptions, giant outputs |
| Memory | Carries facts across turns and sessions | Stale or contradictory notes |
| Control | Budgets, approvals, stop conditions | Infinite loops, silent writes |
Tools are the product
The single biggest lever I've found isn't the model — it's tool design. A good tool:
- does one thing, named like a verb (
publish_site, notsite_ops) - returns small, structured output the model can act on
- fails loudly with an error message that tells the model what to do next
When an agent misbehaves, I read the tool descriptions before I touch the prompt. Nine times out of ten, that's where the ambiguity lives.
Where humans stay in the loop
Reads can be automatic. Writes that are hard to undo should pause for a person. That one rule removes most of the fear from shipping agents.
An agent is only as trustworthy as its least-guarded tool.
Takeaway
Don't start with frameworks. Start with the loop, three well-designed tools and a clear stop condition. You'll be surprised how far that gets you.