Few Tools, Clear Boundaries
Tools are how an agent reaches the outside world. Start with a few high-signal tools, rank them by risk, and draw the boundary before you hand any of them out.
North can read and write my whole notes folder. But it cannot touch a single path outside that folder.
I drew that line before I gave it any tool.
Tools are the agent’s hands
A tool is how the agent reaches the outside world. The talk by 「大大帶我飛」 splits tools into four kinds, from low risk to high risk:
The same LLM with a read-only search tool and with a production database carries a very different risk. The stronger the tools, the more the harness matters, as The Harness Is the Agent’s Workplace explains.
Start with 5 to 10 high-signal tools. Tools that overlap make the agent pick the wrong one from too many similar choices.
Three ways to lose control
North’s boundaries
- It can touch only the notes folder. Every path outside it is blocked.
- Secrets live outside the folder. The program reads them at startup, and the LLM cannot read them.
- Each trigger sends one reminder only, so it never nags me.
- It proposes new goals but never writes them in for me. Any action that changes something stops and waits for my yes.
That last rule is also part of the outer loop in Three Loops in an Agent System: handing off to a human is one way to end a task.
Draw the boundary first, then hand out the tools. If you do it the other way, you patch holes later.
References:
- A talk by 「大大帶我飛」
Related: see why stronger tools need a stronger harness in The Harness Is the Agent’s Workplace, see how a hand-off ends a task in Three Loops in an Agent System, or go back to the series overview: Design an Agent.