It started with a manga, not a paper
I'm a One Piece fan. A while ago I was rereading the Egghead arc, the one with Dr. Vegapunk, the genius who splits his mind into six "satellites". Each one carries a different part of his personality: logic, evil, ideas, knowledge, violence, greed. The original body, Stella, stays in charge. And all of them share a single brain in the cloud called Punk Records.
I stopped on a page and thought: wait, these are agents.
Six specialists with clear roles. A coordinator who decides. A shared memory where everything they learn ends up. I build software for a living, and I've spent the last months playing with multi-agent setups. This was the clearest picture of a multi-agent system I'd seen, and it came from a manga.
But the detail that really got me was another one: the satellites don't need to chat with each other all the time to work together. They sync through Punk Records. That idea became vegapunk-agents, a Multi-Agent Architecture Pattern (and Conceptual Framework). This post explains why my agents don't talk to each other, and why I think that's a feature.
The problem with agents that chat
The most common way to build multi-agent systems today is conversation: agent A sends a message to agent B, B answers, maybe C joins in. It looks natural, because that's how human teams work. In practice, it breaks in predictable ways:
- Loops. Two agents politely disagree, or politely agree, forever. Nobody has the authority to stop.
- Lost context. Each agent only sees its own slice of the conversation. A finding from step 2 is gone by step 7.
- Token bills you can't predict. Every message drags the history along. Costs grow with the number of connections, not with the size of the task.
- No traceability. When the final answer is wrong, you can't point at the moment the reasoning went off the rails.
- No clear owner of the decision. If everyone can talk to everyone, who actually says "done"?
With six agents talking freely, there are fifteen possible channels between them. That's the tangled network you see on the left of the image below.
The idea: a shared blackboard
In vegapunk-agents, satellites never talk to each other directly. Everything goes through Punk Records, a shared memory that works as a blackboard. A satellite reads what's there, does its job, and writes a new entry. The next satellite picks it up from there.
This isn't new. The blackboard architecture comes from AI research in the 1970s and 80s. What's new is applying it to LLM agents, where the cost of free conversation is measured in tokens, loops and hallucinated agreements.
What you get in exchange:
- One source of truth. Every finding, idea, critique and verdict lives in one place, in order.
- Six connections instead of fifteen. Each satellite only connects to Punk Records.
- A full audit trail. You can replay the run entry by entry and see exactly who wrote what, and why.
- Controlled context. Each satellite gets the entries it needs, not a whole chat history.
The rule applies to every implementation of the pattern. Even in CrewAI, where agents can delegate to each other, delegation is turned off.
The cast: Stella and six satellites
Each satellite has a deliberate cognitive bias. That's the point: you don't want six copies of the same balanced assistant, you want specialists that pull in different directions, and a process that turns the tension into a decision.
| Role | Name | Bias | What it does |
|---|---|---|---|
| Orchestrator | Stella | The human | Sets the goal, picks the mode, approves critical actions |
| Memory | Punk Records | None | Shared, append-only record that everyone reads and writes |
| Punk-01 | Shaka | Logic | Checks coherence, rules and safety; the only one who gives the final verdict |
| Punk-02 | Lilith | Criticism | Red team: attacks proposals, looks for flaws and failure points |
| Punk-03 | Edison | Ideation | Unfiltered ideas, brainstorming, first design of the solution |
| Punk-04 | Pythagoras | Research | Finds sources, extracts data, structures it for Punk Records |
| Punk-05 | Atlas | Execution | Heavy lifting: bulk code, assets, batch processing |
| Punk-06 | York | Operations | Background chores: cleanup, formatting, logs, checking deliverables |
Stella is the human in the loop. If an AI coordinator runs the cycle on her behalf, it can move things along, but it never approves or aborts on its own.
The rules that make it work
A blackboard alone isn't enough. Without rules, you just moved the chaos to a shared file. These are the rules the core spec enforces:
- Append-only. Entries are never edited. A correction is a new entry that references the old one. History stays intact.
-
Explicit blocking. Every critique from Lilith must say
blocking: trueorfalse. It's true only if there's a critical or major issue. Implementations never guess it from the text. -
One judge. Only Shaka writes a
verdict. The schema enforces which author can write which entry type. - Shaka can overrule. If a blocking critique isn't grounded in the goal, the constraints or the findings, Shaka can dismiss it, with a written reason.
- Bounded debate. The critique and revision loop runs two iterations by default. At the limit, Shaka must choose: approved or rejected. No endless loops.
- Human before action. If Atlas is going to execute anything, Stella approves first.
You don't always need the full team, so the pattern has three activation modes. Shaka is in all of them:
- Light (2): Edison and Shaka. Quick idea, quick check.
- Standard (4): Pythagoras, Edison, Lilith and Shaka. Research, propose, attack, decide.
- Full Egghead (6): all six satellites, including execution and operations.
The lifecycle
Every run follows the same state machine:
DEFINED → RESEARCH → IDEATION → CRITIQUE ⇄ REVISION → VERDICT → [HITL_APPROVAL] → EXECUTION → OPERATIONS → COMPLETED
Plus two exits: FAILED and ABORTED.
The interesting part is the middle. Edison proposes, Lilith attacks, Edison revises, until Lilith has nothing blocking left or the iteration limit hits. Then Shaka decides. If a phase has no satellite in the current mode, it's skipped: in Light mode there's no research and no critique, just ideation and verdict.
Each transition also emits an event. That stream (events.jsonl) is what will feed the Egghead Monitor, a live visualizer where you'll see the satellites light up as they write to Punk Records.
A pattern, not another library
There are plenty of multi-agent frameworks already. I didn't want to build a new one and ask you to switch. vegapunk-agents is a specification first.
The core/ folder is the single source of truth:
-
architecture.md: the protocol, written with RFC 2119 keywords (MUST, SHOULD, MAY). -
punk-records.mdplus JSON Schemas for entries and events, with examples. -
prompts/: canonical system prompts for Stella and the six satellites, sharing one common protocol. - A validator that runs in CI on every change.
On top of that come the flavors: implementations of the same spec on different engines. Claude Code native subagents, LangGraph, CrewAI and the plain Anthropic SDK. They can use each engine's strengths, like native structured output, but they can't redefine the rules. Demos live apart in examples/, so the framework never depends on a use case.
If you already use one of those engines, the idea is that you can adopt the pattern without leaving it.
What's next
The core spec is out as v0.1.0. From here, the roadmap goes flavor by flavor:
-
Claude Code native: each satellite as a subagent, generated from
core/. - LangGraph + first demo: a real Edison ⇄ Lilith → Shaka loop, plus the Egghead Startup Validator, which uses free DuckDuckGo search, no extra API keys.
- Egghead Monitor: the live visualizer.
- CrewAI and Anthropic SDK: same demo, same behavior.
I'll write about each step in this series.
If the idea clicks with you, take a look at the repo, open an issue, or try the spec with your own agents. I'd love to hear where it breaks.
- Repo: github.com/ckomiya/vegapunk-agents (Apache-2.0)
- Cite it: doi.org/10.5281/zenodo.24384581
vegapunk-agents is inspired by Dr. Vegapunk from One Piece, created by Eiichiro Oda and published by Shueisha. It's a fan-inspired open-source project with no commercial affiliation.




Top comments (0)