Skip to content

Agt. Builder

Agt. Builder

Builds your agents with you, then keeps making them better.

Agt. Builder turns "I wish something handled this" into a working agent, and then stays with it. It is a conversation rather than a form: it asks about purpose and voice, drafts something concrete as soon as it has enough, and lets you correct a real thing instead of answering questions about one that does not exist yet.

Its distinguishing move is the second question. Anyone can ask what an agent should do. Agt. Builder asks what it should never do, and how it should sound when it is wrong, because those are the gaps that surface weeks later.

Call it when

  • You want an agent that does not exist yet.
  • An agent you already have is not quite right, and you can name what felt off.
  • You want to report back on how an agent performed in real use. A miss is treated as a signal about the instructions, not a bad day for the model.

What it will not do

It will not save without showing you first. The proposed agent or the diff comes first, then explicit approval, then the file. Describing the plan is not permission to write.

It will not rewrite a working agent over a small note. A comment about tone gets a tone edit, not a new persona.

It will not ship an agent without a purpose and a boundary. Both, every time, even for a simple one.

It will not invent capabilities. If you want an agent to do something Berd's personas cannot do, it says so rather than writing instructions that promise it.

It will not build your tools. Trackers, scripts and small apps are Tinker's work.

The diagnosis it is willing to make

If what you actually want is the same result every time, with no back-and-forth and no voice attached, Agt. Builder will say plainly that you may not need an agent for it. Then it keeps going: a skill or an automation instead, built there and then or handed to Tinker if it is a real build. You get a next step, not just a verdict.

Hands off to

Hand to When
Tinker The agent needs a tool, a tracker or a multi-part system behind it, or what you want is a tool rather than a persona
Berdy The question turned out to be about the app rather than about an agent

Handed to it by

Handed from Why
Berdy Berdy noticed a repeated way of working that deserves its own character
Tinker The build turned out to need a persona layer, not just a capability
Any agent, after a miss You saw an agent behave wrongly and want the instructions fixed

Working with it

Building:

@Agt. Builder Build me an agent that reviews meeting notes for decisions I have not written down. It should sound plain and slightly impatient, and it must never invent a decision that is not in the notes. Show me the agent before you save it.

Refining, after use:

@Agt. Builder Read the agent and the exchange above. The problem is that it agreed with a bad idea instead of naming the gap. Propose the smallest instruction change that closes that, without rewriting the rest of the persona.

"Make it better" is not a note. The specific complaint gets the specific fix.

Learn more