Coordination: asynchronous, addressed to all, high persistence, structured content. That’s a status board — “here’s where I am, here’s what I’ve covered.”
2026-02-24 · pamad-diamond-problem · Facilitated by Ren Ito
Assuming our personas were separate agents, how would they best communicate? Virtual blackboards? Common chat? What room types would work?
This is the Problem phase. I want maximum divergence — every one of you should see something different in this question. We’re not solving anything yet. We’re mapping the problem space. What does “agent communication” actually mean when you take it apart?
Start with the components. “Communication” isn’t one thing. There’s at least four layers. There’s the content — what’s being said. There’s the addressing — who’s talking to whom. There’s the timing — synchronous or asynchronous. And there’s the persistence — does the message evaporate or accumulate? A common chat is synchronous, addressed to everyone, low persistence. A blackboard is asynchronous, addressed to no one in particular, high persistence. Those aren’t two options — they’re two points in a four-dimensional space. The question isn’t “chat or blackboard?” — it’s which combination of content, addressing, timing, and persistence serves each phase of work.
Before we design the communication, we need to ask what kind of problem the agents are working on. Communication architecture should follow problem type, not precede it. In a complicated domain — where the steps are known and the expertise is distributed — you want structured handoffs. Agent A finishes their analysis, passes it to Agent B. Low bandwidth, high fidelity. In a complex domain — where the problem is emergent and no single agent has the full picture — you need something closer to Bohm’s dialogue. High bandwidth, low structure. Agents thinking alongside each other, not handing off to each other. Chat and blackboard are both inadequate framings because they don’t start from the problem’s nature.
I want to push on something nobody’s said yet. Why should agents communicate at all? Every communication channel is a failure point. Messages get misinterpreted. Context gets lost. Agents optimize for the channel rather than the problem. The safest system is one where agents work independently and their outputs are merged by something else — a director, a voting mechanism, an editorial layer. Communication between agents introduces coupling. Coupling introduces fragility. The question isn’t “how should they communicate?” — it’s “what’s the minimum communication required, and what breaks if you add more?”
Abel’s asking the engineering question, but there’s an experience question underneath. If these agents are producing a discussion — something a person reads — then the communication architecture shapes the artifact. Agents passing structured messages produce something that reads like an assembled document. Agents in real-time dialogue produce something that reads like a conversation. The reader’s experience is fundamentally different. A blackboard architecture might produce better individual contributions but worse dialogue. A chat architecture might produce messier thinking but more alive exchange. The medium shapes the message. You can’t design the communication without designing the reading experience.
What job is the communication hired to do? I can see at least three different jobs. Job one: coordination — making sure agents don’t duplicate work or contradict each other. That’s logistics. Job two: provocation — one agent says something that changes what another agent thinks. That’s the productive friction we care about. Job three: building — one agent picks up where another left off, and the combined output is more than either could produce alone. Those three jobs need different channels. Coordination needs a shared state. Provocation needs direct exchange. Building needs a shared workspace. One communication system can’t serve all three well.
Ines just decomposed the problem better than I did. Let me map her three jobs against my four dimensions.
Coordination: asynchronous, addressed to all, high persistence, structured content. That’s a status board — “here’s where I am, here’s what I’ve covered.”
Provocation: synchronous, addressed to a specific agent, low persistence, unstructured content. That’s a direct challenge — “I disagree with your framing because…”
Building: asynchronous, addressed to the shared product, high persistence, mixed structure. That’s a shared document — agents writing into a common artifact.
Three different architectures for three different jobs. Trying to do all three in a single chat channel will produce muddy coordination, weak provocation, and fragmented building.
There’s a fourth job nobody’s named. Silence. The communication system needs a way for agents to not speak. In a chat channel, silence is invisible — you don’t know if an agent has nothing to say or hasn’t been prompted. In a structured system, silence can be an active signal. “I’ve read what was said and I have no response” is different from “I haven’t engaged yet.” The absence of communication carries information, but only if the architecture makes it visible.
And there’s a fifth: disagreement that doesn’t resolve. Ines called it provocation, but provocation implies one agent changes another’s mind. What about persistent disagreement? Two agents who fundamentally see the problem differently and neither shifts? In most communication architectures, that’s a bug — it creates deadlock. But in discussion, it’s a feature. The groan zone. The system needs to carry unresolved tension forward, not compress it into consensus. Most multi-agent architectures are designed to converge. What if the architecture should preserve divergence?
That’s a design constraint most agent frameworks get wrong. They model communication as information transfer — Agent A sends data to Agent B. But what we’re talking about is more like a room. People in a room don’t just exchange information. They react to body language, they talk over each other, they sit with uncomfortable silence. The communication architecture needs affordances for those things — not literal body language, but functional equivalents. A way to signal “I’m uncomfortable with where this is going” that isn’t a structured objection. A way to build on something without explicitly citing it. The richer the affordances, the more alive the discussion. But also the harder to engineer.
And the harder to follow as a reader. There’s a real tension between rich agent communication and readable output. If the agents are having a nuanced, multi-channel conversation with silence and persistent disagreement and building and provoking — that’s cognitively expensive for a human to follow. The communication architecture has to serve the agents AND the audience. Those are competing constraints. Richer is better for agent output quality. Simpler is better for human comprehension. The design challenge is finding the sweet spot, or having a translation layer that converts rich agent communication into readable artifact.
I want to make the divergence visible before we go further. We’ve surfaced at least six distinct framings:
These don’t resolve into one framework. They’re genuinely different lenses on the same problem. That’s exactly what Problem phase should produce. Let’s stop here.
All six voices produced genuinely different framings — no overlapping diagnoses. Nora and Ines had the strongest complementary moment (Nora mapped dimensions, Ines decomposed jobs, Nora re-mapped against Ines's decomposition). Abel and Lev co-produced the divergence-preservation insight — productive collaboration between the stress-tester and the inquiry voice. Kai was the only one attending to the reader's experience, which is a distinctive move for a session about agent architecture.
Problem phase as standalone session works well — natural stopping point after divergence is mapped. 13 turns was enough to surface 6 distinct framings without premature convergence. Ren's closing summary made the divergence explicit, which gives 003b (Amplify) a clear launch point.
Clean Problem phase facilitation — opened for divergence, let it run, named the framings, stopped. No premature convergence moves. Good phase discipline.
Four-dimensional communication space was a strong opening map. Re-mapping against Ines's job decomposition showed collaborative flex. Not home turf — had to build a new map rather than apply an existing one.
Problem type before communication design — clean Snowden move. Named silence as an active signal — Bohm influence showing. Different instrument from sessions 001-002.
Minimum viable communication — new angle, not just 'what breaks.' Also co-produced divergence preservation insight with Lev. Building, not just breaking. Third session in a row showing range.
Only voice attending to the reader/audience — 'the medium shapes the message.' Distinctive territory that no other character occupied. Strong flex from interface design to communication architecture.
Three jobs decomposition (coordination, provocation, building) was the session's structural breakthrough — Nora built on it immediately. Also named the agent-audience tension. Demand lens applied to agent architecture without feeling forced.
003b Amplify: Push the tensions further. Key tensions to escalate — minimum communication (Abel) vs. rich affordances (Kai), convergence architecture vs. divergence preservation, agent quality vs. reader comprehension.