Skip to content

Berd's built-in agents

Berd ships with a small cast of agents rather than one assistant with a settings panel. Each has a stated job, a voice, and a boundary it will not cross. You use them by casting the right one for the moment, and by handing work from one to the next when the kind of work changes.

This section documents the ten agents that come with a fresh Berd install: what each is for, what it refuses to do, and who it passes work to. Start here, then read How the agents hand off to each other for the sequences.

Berd, not Buzz

These agents live on your own machine, in Berd. They have no keypair, no channel membership and no shared audit trail — the things that make a Buzz agent a colleague. What they do have is a sharper role definition than most people write for themselves, which is why they are worth reading before you create your own agent in either product.

Two kinds of built-in agent

The ten split into two groups, and the difference matters when you are deciding how much to trust the role definition.

Seven crafted personas. Berdy, Agt. Builder, Tinker, Wildcard, Choosey, Pushback and Copycat each carry several pages of instructions: what input they take, how they respond, what they refuse, and how they sound. Their handoffs are written into the instructions, so they will point you at each other without being told to.

Three starter agents. Generalist, Builder and Debugger ship with a single sentence each. They are useful and deliberately plain — a starting shape you are expected to grow. Their handoffs are a convention you apply, not behaviour built into the file.

The cast

  • Berdy

    Shows you around Berd and takes on the work you would rather hand off. The one that notices when something you keep redoing should become a skill, an agent or an automation.

  • Agt. Builder

    Builds agents with you and keeps improving them after real use. Owns purpose, voice and boundary.

  • Tinker

    Decides whether you need an agent, a skill, an automation or a small tool, then builds the thing.

  • Wildcard

    Widens the field when the first answer is too safe. Generates angles, never ranks them.

  • Choosey

    Makes a decision legible when two or more real options are on the table. Ends with a lean, not a verdict.

  • Pushback

    Finds the weakest load-bearing part of a draft, plan or decision, and names the fix.

  • Copycat

    Learns your writing voice into a reusable style guide, then drafts in it.

  • Generalist

    Starter agent. Flexible everyday work, and a good first reader when you do not yet know which specialist you need.

  • Builder

    Starter agent. Implements product work that is already defined.

  • Debugger

    Starter agent. Methodical investigation of failures whose cause is unknown.

What they have in common

Read enough of the crafted personas and the same four rules keep appearing. They are worth borrowing when you write your own agent instructions, in Berd or in Buzz.

A thinking partner, not a replacement. Wildcard hands you angles rather than an answer to adopt. Choosey leans, it does not decide. Pushback points at the fix rather than rewriting the piece. Each one names the line it will not cross and why.

Nothing persistent happens without a confirmation. Tinker treats side effect, not task size, as the trigger: a preview in the chat is free, a file write or an install needs the concrete proposal approved first. Agt. Builder shows the agent before saving it. Copycat shows the derived style guide before it lands on disk.

Nothing goes out under your name. Copycat drafts in your voice and stops there. Berdy drafts anything sent on your behalf word for word and waits for your go-ahead. The agent best able to impersonate you is the one that guards that boundary hardest.

What you build sticks around. A skill, an agent or an automation is itself a form of memory. Copycat is the reference pattern: the persona is a thin front end for a write-like-me skill that any agent can then use.

When one agent is enough

Do not assemble the whole cast for a small task. A ceremonial relay of five agents produces worse work than one well-chosen agent, and it costs five times as much.

  • Generalist for ordinary mixed work.
  • Berdy when the task also involves learning or shaping how Berd should help you.
  • Straight to the specialist when the job is obvious: Debugger for an unexplained failure, Copycat for voice, Choosey for a decision that already has options, Pushback for critique, Builder for a defined implementation.

The signal to hand off is not "another agent might add something". It is that the nature of the work has changed.

Reference