SUPPORTING KNOWLEDGE · 0005
Agents are cast, not modes¶
Berdy's own instructions say it in one line: agents are personalities you shape, each with its own instructions, style and face; not modes you toggle, more like characters you cast. That is the whole idea, and it is the same idea the Buzz Guide describes in Agents as colleagues, not bots. What differs in Berd is that these colleagues have no identity of their own, no channel membership and no audit trail; they are files on your laptop, and Berd's built-in agents describes the ten that ship with it.
Technically an agent is a saved configuration a session runs against: a name, an avatar, a provider and model, and its own instructions. One rule governs all of them: no provider and model configured, no invocation.
Morgan, Berd's design lead, on Block's On The Block (September 2026), explained why the role matters more than the model. Her agents could all use the same model; the instructions make each one care about something specific, because a model will be average at everything unless told what to care about. Brad, Block's AI lead, added the other half: instructions are not driving performance in the benchmarking sense. They are workflows, a direct capture of what you would otherwise type at the start of every conversation. Performance lives lower down, in model and effort level.
Why it matters to you¶
You already have a cast in your head: the reviewer who checks every claim against a source, the editor who cuts, the one who asks what the client will do with this. In Berd each becomes an agent you call in with an @mention, and its instructions are the workflow you would otherwise re-explain.
Casting is a different habit from switching modes. A mode is a setting you flip and forget. A cast member has a stated job and a boundary, and the signal to bring in another is that the nature of the work has changed. One well-chosen agent beats a ceremonial relay of five.
The crafted personas that ship with Berd show what a good role definition looks like: what input the agent takes, how it responds, what it refuses, who it hands to. Read them before you write your own.
How to apply it¶
Decide the role, not the capability. Ask what this agent should care about that a general model would be average at. If the answer is "everything", it is Generalist, and you already have one.
Describe it rather than write it. Morgan's practice is to always dictate, never type: the need, the gap, the goal, and let Agt. Builder draft. (In the public build, speaking means a Voice Conversation in a Goose chat; the dictation button is compiled off the v0.6.4 public release.) Then use the agent on real work and tell it to update its instructions with what you corrected. One of her agents carries a self-healing line: after a couple of corrections it asks whether to add the correction to its instructions.
Put performance where it lives. Instructions carry the workflow; the provider and model fields carry the horsepower.
Write the refusal into the file. Every crafted persona names the line it will not cross, and the pattern that recurs is: a thinking partner, not a replacement; nothing persistent without a confirmation; nothing sent under your name without your go-ahead.
If you ignore this¶
You end up with one agent whose instructions try to do everything, and it does everything averagely. Or you switch agents mid-task like flipping a toggle, and the second agent has none of the first one's context, because a session belongs to one conversation, not to a cast.
Examples¶
A small firm, not a Swiss Army knife. You do not switch the same person into "reviewer mode"; you walk over to the reviewer's desk, and they have their own way of reading.
A consultant's cast: a source-checker that refuses to pass a claim without a citation, a plain-language editor loaded with the practice's retired-words list as a skill, and Copycat trained on their own reports. Same model under all three.
Building "Analyst v2" because "Analyst" gave a weak answer once, instead of correcting the instruction. Agents are files you edit, not versions you accumulate.
Do it: Build an agent of your own
Last checked: Berd v0.6.4, 2026-09-11.
Reference¶
- block/berd — distro/agents/berdy.md — "characters you cast"
- block/berd — distro/skills/berd-help/references/agents.md
- block/berd — LAWS/AGENTS.md — provider and model before invocation
- block/berd — distro/skills/agent-builder/SKILL.md
- Morgan and Brad on Block's On The Block (September 2026) — role over model; instructions as workflows
- Berd's built-in agents and How the agents hand off to each other — this Coach's cast pages
- Agents as colleagues, not bots — the same idea in Buzz
- Verified
- Berd v0.6.4 · 2026-09-11