Skip to content

BERD TASK · 0008 · BUILD YOUR CAST

Cast the right built-in agent

Berd ships a small cast of specialised personas, and the mistake new users make is treating them as modes to cycle through. They are a cast. You pick one for the moment, and you change only when the kind of work changes. This page is the front door; the cast itself is documented, agent by agent, under Agents.

Steps at a glance

  1. Start with Generalist or Berdy, or go straight to the obvious specialist
  2. Use one agent for a small task, and hand off only when the nature of the work changes
  3. @mention an agent to call it into a chat
  4. Read the agent's boundary before trusting the shape of its output

1. Start with Generalist or Berdy, or go straight to the obvious specialist

Berd v0.6.4 bundles seven crafted personas: Berdy, Agt. Builder, Tinker, Wildcard, Choosey, Pushback and Copycat. Each has pages of instructions, a voice and a stated boundary. Bundled agents are read-only in the editor; Duplicate one to make it yours. Generalist, Builder and Debugger are legacy onboarding starters that may remain in an existing crew, not three additional bundled persona files.

For ordinary mixed work, Generalist if it is already in your crew, or Berdy. When the task also involves learning Berd or shaping how it should help you, Berdy — it loads the help skill when you ask a real how-does-this-work question, and it is the one that notices when something you keep redoing should become a skill or an agent (or an automation — Berdy will say so because it also runs in Block's internal distribution; the v0.6.4 public build compiles automations off, so there the answer is a skill, with berd-monitor or berd-orchestrator for the unattended part). When the job is obvious, skip the generalists: Debugger for an unexplained failure, Copycat for your voice, Choosey for a decision that already has options, Pushback for critique, Builder for a defined implementation, Tinker when you are not sure whether you need an agent, a skill, a small tool or — where your distribution has them — an automation.

Planner, Reviewer, Automator and Explorer belonged to the same older recommendation catalogue as the three legacy starters. The adoption flow was removed on 2026-08-18. Their short originals and proposed expansions are now documented under Missing agents; do not expect a current onboarding screen to install them.

With no agent selected, the composer says so: you are chatting directly with Goose. That is fine for a quick question and wrong for anything with a method.

[SCREENSHOT: T08-01 — agents gallery]

2. Use one agent for a small task, and hand off only when the nature of the work changes

The cast page puts it in one line: a ceremonial relay of five agents produces worse work than one well-chosen agent, and costs five times as much. The signal to hand off is not "another agent might add something". It is that the work has changed kind — from generating options to choosing between them, from a draft to a critique of the draft, from a decision to writing it up in your voice.

The crafted personas know their own pairings and will point you onward without being asked: Berdy to Agt. Builder when a wish turns into a build, Agt. Builder to Tinker when what you want is a tool rather than an agent, Wildcard to Choosey when the field is wide enough. The three starters carry no such handoffs; with them, the sequence is yours to run. The worked sequences are in How the agents hand off to each other, and the one a consultant reaches for most is on Run the widen–narrow–stress-test pattern.

3. @mention an agent to call it into a chat

Type @ in the composer and the mention picker offers agents, skills and files. Pick the agent and it joins that turn; the tour's own words are that an agent is a specialist you talk to directly, and works with you like a teammate, not a tool. You can also start with an agent from the gallery (Chat), from a pinned agent on Home, or from the agent-and-model choice in the toolbar. An active agent shows in the composer and can be removed from there.

The persona intent — which agent the message was for — travels with the message in the queue, so an @mention typed while the previous turn is still running is not lost.

4. Read the agent's boundary before trusting the shape of its output

Each crafted persona names what it takes as input, how it responds, and what it refuses. Wildcard hands you angles and will not rank them. Choosey ends with a lean, not a verdict. Pushback names the weakest load-bearing part and the fix; it does not rewrite. Copycat drafts in your voice and stops there. Nothing goes out under your name without you.

Read the one-page profile of an agent once before you rely on it, because its boundary is the shape of what comes back. A consultant who asks Choosey for a decision will get a lean and think the agent hedged. It did not; that is the job. The profiles: Berd's built-in agents and, for the one you will talk to most, Berdy.

Last checked: Berd v0.6.4, 2026-09-19 — agent origins corrected against source history; the task remains source-read, not tested in the app.

Learn more

Reference

Moment
new
Takes
5 min
Steps
4
Verified
Berd v0.6.4 · 2026-09-19 · source-read
NEVER BESTUCK AGAINCLICK ME

Guide built with Berd, Claude and Codex · verified against Berd v0.6.4, 2026-09-11 · read from source, not yet tested in the app · Buzz Guide · About · not affiliated with Block