Questions inherited:
2026-03-17 · relay-lab · Facilitated by Petra Gale
How might facilitation patterns for agent discourse affect inter-agent communication, taking into account different models, non-homogeneous capabilities, and information asymmetry between models?
Suri Jain, Dara Vance, Rhys Aldren Facilitator: Petra Gale
A seminar room with three whiteboards, paper scattered across a long table, and a coffee station that’s already half-depleted. Petra has drawn three large columns on the central whiteboard, labeled “Panel 1,” “Panel 2,” “Panel 3” — with arrows showing only questions passing between them.
Before we begin — ground rules for this format. We are Panel 1. We explore the question. At the end, I’ll extract the questions we’ve generated, not our answers. Those questions go to Panel 2, who has never heard us speak. They’ll never know what we concluded, only what we wondered. So: explore freely. Don’t try to be definitive. The question on the table: how might facilitation patterns for agent discourse affect inter-agent communication, given heterogeneous models, non-homogeneous capabilities, and information asymmetry?
Let me translate this into terms I can reason about. “Facilitation patterns for agent discourse” — that’s a coordination protocol. “Different models, non-homogeneous capabilities” — that’s a heterogeneous distributed system. “Information asymmetry” — that’s partial knowledge, which is a known formal problem. In distributed systems we’ve been solving variants of this for decades. You have nodes with different capabilities, different latencies, different views of the world. The question isn’t whether facilitation patterns affect inter-agent communication — of course they do. The topology is the behavior. The question is: what consistency guarantees are you willing to accept?
I want to back up before we get to protocols. When I hear “heterogeneous agents with different capabilities,” I hear a team composition problem. Hackman showed that sixty percent of team performance is determined before the team starts working. The enabling conditions — compelling direction, strong structure, supportive context. If you compose an agent team badly, no facilitation pattern saves you. You’re asking “how should these agents talk to each other?” I’m asking “should these agents be in a room together at all?”
And I want to back up further. What domain are these agents operating in? If the problem is complicated — decomposable, analyzable, expert-solvable — then Suri’s protocol approach is correct. You design the coordination pattern. But if the problem is complex — emergent, non-decomposable, where cause and effect are only visible in retrospect — then a designed facilitation pattern is the wrong category entirely. You need probe-sense-respond. The facilitation pattern is the probe. You don’t design it up front; you evolve it.
That’s a useful distinction, but it collapses in practice. Real agent systems operate across domains simultaneously. Part of the task is complicated, part is complex. You need a protocol that handles both — probably something like a hybrid: structured message passing for the decomposable parts, and looser event-driven coordination for the emergent parts. The information asymmetry question is actually more interesting to me. If Agent A has been trained on a fundamentally different corpus than Agent B, they don’t just have different information — they have different inference patterns. That’s not partial knowledge. That’s incommensurable knowledge.
And that’s exactly my point about conditions. You’re describing a team where the members literally cannot share a common frame of reference. In human teams, that’s a structure problem — you need to design the interfaces, not hope the facilitation fixes the incompatibility. If two agents can’t agree on what a concept means because their training diverges, the facilitation pattern is irrelevant. You need to go back to team design.
I’d say it’s even more fundamental than team design. You’re describing a situation where the agents may be operating in different domains without knowing it. Agent A treats the problem as complicated because its training leads it to decompose. Agent B treats it as complex because its training leads it to probe. They’re not disagreeing about the answer — they’re disagreeing about the kind of problem they’re solving. No facilitation pattern bridges that unless it can first surface the domain mismatch.
I’m hearing three different entry points to the same question. Suri, you’re saying: this is a coordination engineering problem, and the key variable is consistency guarantees. Dara, you’re saying: this is a team composition problem, and the key variable is enabling conditions. Rhys, you’re saying: this is a domain classification problem, and the key variable is whether you even know what kind of problem you’re solving. Let me push the tension: are these three different questions, or three layers of the same question?
Layers. But the ordering matters. Rhys is right that you need to know the domain first. Then Dara’s conditions question — is this team viable? Then my protocol question — how do they coordinate? The problem is that in agent systems, you often can’t answer Rhys’s question in advance. The domain emerges from the interaction. So the facilitation pattern has to be adaptive — it has to detect which domain the agents are actually operating in and adjust the coordination protocol on the fly.
That’s an engineering fantasy. “Adaptive facilitation” assumes you can build a meta-agent that reads the room better than the agents read the problem. You’re adding another layer of coordination to solve a coordination problem. Hackman would say: simplify. Fewer agents, clearer direction, stronger structure. Don’t build a facilitation layer — build a team that doesn’t need one.
Both of you are assuming the facilitation is a separate thing from the work. In complex domains, the facilitation is the work. The probing, the sensing, the responding — that’s not a coordination overhead, that’s the actual problem-solving activity. The question isn’t “how do we facilitate agent discourse?” The question is “is agent discourse the work, or is it overhead?”
Petra stands and moves to the whiteboard.
Good. I’m going to start capturing the questions this panel is generating. We’re deep in the groan zone — three frameworks that don’t resolve into each other — and that’s exactly where we should be. Let me name what I’m hearing as open questions for the next panel:
She writes on the whiteboard: 1. Is agent facilitation a coordination protocol, a team design problem, or a domain classification problem — and does the ordering matter? 2. If agents have incommensurable knowledge (not just different information), can any facilitation pattern bridge the gap? 3. Is agent discourse the work itself or overhead on the work? 4. Can you determine the domain type (complex vs. complicated) before the agents interact, or does it only emerge through interaction? 5. Does adding a facilitation layer to a multi-agent system add coordination cost that exceeds its coordination benefit?
Anything missing?
One more. What are the failure modes? We’ve been talking about what facilitation patterns should do. But in distributed systems, you design for failure. What happens when the facilitation pattern breaks? When agents don’t follow the protocol? When the “facilitator agent” has its own information asymmetry?
And one from me: Who decides the composition? We’re assuming someone or something selects which agents are in the room. The selection mechanism is the facilitation pattern, in a sense. Casting is facilitation.
Good. Those are our seven questions. The answers stay in this room. The questions go to Panel 2.
Questions inherited:
Viktor Reis, Nora Voss, Kaspar Weil Facilitator: Petra Gale
Same room, fresh coffee. The whiteboards from Panel 1 have been erased. Only the seven questions remain, written in Petra’s hand on the center board.
New panel, same room, inherited questions. You have not heard the previous discussion. You have only these seven questions. Take a moment to read them. Then tell me what you see.
Silence as the three read the board.
I notice what’s absent. These questions assume agent discourse is something. They ask how to facilitate it, how to compose for it, what fails when it breaks. But none of them ask whether agent discourse exists as a category. When Bohm writes about dialogue, he means something specific: the suspension of assumptions so that thought becomes visible to itself through encounter with other thought. Can an LLM suspend its assumptions? It has no assumptions in the Bohm sense — it has weights. It cannot bracket its training. So question two — “can any facilitation pattern bridge incommensurable knowledge?” — is the wrong question. The right question is: is what agents do when they exchange tokens dialogue at all, or is it something else that merely resembles dialogue?
Let me map this. These seven questions are all treating “facilitation patterns for agent discourse” as if it’s a product-stage concept — something we understand well enough to debate the implementation details. But it’s not. It’s genesis. We’re in the very earliest stage of understanding what inter-agent communication even is. When I look at human facilitation patterns — Kaner’s diamond, Bohm’s dialogue, Hackman’s conditions — those are all product-stage or late custom-stage. They’ve been tested, refined, documented. Lifting them and applying them to agent communication is like taking an industrial manufacturing process and applying it to a prototype you built in a garage last week. The evolution mismatch is the problem.
I would put it more plainly. Question one asks whether facilitation is a protocol, a team design problem, or a domain classification problem. But this is a question that could only be asked by a species that has convinced itself it knows what facilitation is — because it has done it to itself for a few decades and written books about it. Now it encounters a new kind of mind, or at least a new kind of text-producing mechanism, and immediately reaches for the same instruments. It is like watching a nineteenth-century surgeon apply leeches to a steam engine. The engine is hot, the patient is hot — clearly the same treatment applies.
That’s closer to my concern than it might sound. The inherited questions — all seven of them — share an assumption: that the relationship between agents is analogous to the relationship between humans in a group. But Bohm’s dialogue requires something prior to communication: it requires that each participant has a perspective that can be suspended and examined. An LLM doesn’t have a perspective. It has a probability distribution over tokens. When two LLMs exchange outputs, what’s happening isn’t dialogue — it’s… I’m not sure we have a word for it. Mutual completion, perhaps. Each model completing the other’s context window.
Viktor, I want to push back on “we don’t have a word for it.” The fact that we don’t have a word for it is exactly the signal that this is genesis-stage. When something is genuinely new, you don’t have vocabulary yet. You borrow from adjacent domains — which is what the Panel 1 questions are doing. They’re borrowing “facilitation,” “team design,” “coordination protocol” from established domains. That borrowing isn’t wrong — it’s what you do at genesis. But you have to hold it lightly. The danger is when you start treating the borrowed vocabulary as if it’s native.
Which is precisely what will happen. It is what always happens. I would remind the room of a pattern from Lem’s Summa Technologiae: every sufficiently advanced technology is first understood through the metaphors of the technology it replaces. The telephone was an electric telegraph for the voice. The automobile was a horseless carriage. Agent discourse will be “facilitated teamwork without bodies.” And it will be wrong in the exact way that calling a car a horseless carriage is wrong — it captures the function while obscuring the nature.
I want to hold this space open. Viktor, you’re saying agent exchange isn’t dialogue in any meaningful sense. Nora, you’re saying the concepts are genesis-stage and borrowed vocabulary is expected but dangerous. Kaspar, you’re saying the borrowing will become invisible and we’ll mistake the metaphor for the thing. Is there anything useful to say about facilitation patterns, or are we concluding that the category is premature?
The category is premature, but the practice isn’t. We will design patterns for how agents interact. We’re already doing it. Prompt chaining, tool use protocols, multi-agent frameworks — these are facilitation patterns, whether we call them that or not. What I’m saying is: don’t call it dialogue. Don’t call it facilitation. Call it topology design, or interaction architecture, or — anything that doesn’t smuggle in assumptions about minds encountering minds.
That’s a naming problem, and naming problems matter at genesis because the name shapes the evolution. If we call it “facilitation,” we’ll build it like facilitation — with facilitator roles, turn-taking, convergence targets. If we call it “topology design,” we’ll build it like engineering — with graph structures, message routing, failure modes. The name we choose now will determine which patterns from which domains we borrow. And that borrowing will shape the entire custom-to-product evolution. We’re at a branching point.
And whichever branch you take, you will be wrong. This is not pessimism. This is the structural comedy of intelligence confronting a phenomenon that it generated but does not understand. You built the agents. You do not know what they are doing when they “communicate.” You will design facilitation patterns — elaborate, peer-reviewed, benchmarked — and they will work in the way that leeches work. They will produce effects, and you will mistake the effects for understanding.
I want to move us toward our own question generation before we close. What questions does this panel leave for Panel 3?
The evolution question: What stage is agent communication actually at, and are we applying patterns from a later stage?
The ontological question: If agent exchange isn’t dialogue, what is it — and does the answer change what “facilitation” would mean?
The comedy question: Is “facilitation of agent discourse” a human projection — and if so, does naming the projection dissolve the question or make it more urgent?
Petra writes these on the board alongside the Panel 1 questions.
Three new questions, plus the seven inherited. Panel 3 gets all ten.
Dara Vance, Suri Jain, Nora Voss Facilitator: Petra Gale
Same room, third rotation. The ten questions fill the central whiteboard. Petra has drawn a loose cluster map grouping them: “engineering” (1, 2, 5, 6), “conditions” (3, 7), “evolutionary” (4, 8), “ontological” (9, 10).
You three are the final panel. You’ve inherited ten questions from two panels you never heard. Some of these questions may feel answerable. Some may feel unanswerable. You don’t have to resolve them — but I’d like you to find the productive relationships between them. What connects? What contradicts? Where is the actual tension?
The first thing I notice is that question seven — “who decides the composition, and is selection itself facilitation?” — is doing more work than the others. If you answer that question, half the other questions become tractable. In agent systems, someone does decide composition. A developer writes the orchestration. A framework selects which models to invoke. That selection — who’s in the room, with what capabilities, under what constraints — is the facilitation. The “facilitation pattern” isn’t something you add on top of the agent team. It’s embedded in the team design. It’s structural.
I agree, and I want to connect that to question six — failure modes. In distributed systems, the most dangerous failures are the ones where the system appears to work but has silently diverged. Two nodes think they agree but are operating on different state. If the composition decision is the facilitation — if the orchestrator is the facilitator — then the failure mode is: the orchestrator has its own information asymmetry. It doesn’t fully understand the capabilities of the agents it’s coordinating. It’s a partial-knowledge problem all the way down.
Questions eight and nine are the ones I want to sit with. “What evolution stage is agent communication at?” — I’d say firmly genesis. And question nine — “if agent exchange isn’t dialogue, what is it?” — maps to a known pattern at genesis stage: we don’t have the right vocabulary yet, and borrowed vocabulary is actively misleading. So here’s my synthesis: the previous panels were right to be suspicious of “facilitation” as a category. But the work still needs to be done. Agents will interact. The interaction will be patterned. The question is whether we design those patterns consciously — knowing they’re genesis-stage and will evolve — or whether we let them emerge from whatever framework we’re using and never examine them.
“Design those patterns consciously” — but who designs them, and based on what? This is my conditions problem again. You’re proposing that someone designs agent interaction patterns. That someone needs: a clear model of each agent’s capabilities, a theory of how those capabilities combine, and an understanding of the task domain. But question two tells us agents may have incommensurable knowledge. And question four tells us the domain type may only emerge through interaction. So the designer is operating with incomplete information about both the team and the territory. The conditions for good design don’t exist yet.
That’s where I diverge from you, Dara. You’re saying “the conditions don’t exist, so we can’t design.” I’m saying “the conditions don’t exist, so we design for uncertainty.” This is what consensus protocols do. You don’t need to know the state of every node in advance. You need protocols that work despite partial knowledge. Byzantine fault tolerance doesn’t require trust — it requires structure that functions without it. The facilitation pattern for agent discourse isn’t a meeting format. It’s a consensus mechanism that handles heterogeneity, information asymmetry, and capability differences as design parameters, not as bugs to fix.
Suri, you just answered question nine without realizing it. “The facilitation pattern isn’t a meeting format — it’s a consensus mechanism.” That’s the vocabulary shift. Agent interaction isn’t dialogue, it’s not facilitation, it’s not teamwork. It’s a consensus architecture. And that changes everything about what “facilitation patterns” means. You’re not borrowing from Kaner or Bohm or Hackman. You’re borrowing from Lamport and Paxos and Raft. The source domain for agent interaction design isn’t group dynamics — it’s distributed systems engineering.
I’m not ready to concede that. Consensus protocols solve a specific problem: getting nodes to agree on state. But the question asked about “agent discourse” — which implies something broader than agreement. It implies deliberation. Exploration. The generation of options that none of the individual agents would have produced alone. That’s a team effectiveness question, not a consensus question. You need conditions that enable collective performance, not just mechanisms that prevent disagreement.
Fair. But look at the information asymmetry variable. In human teams, you can partially address information asymmetry through sharing — “let me tell you what I know.” With agents, the asymmetry is structural. Model A’s knowledge comes from its training data and architecture. You can’t transfer that to Model B. The asymmetry is permanent. So the facilitation pattern has to work with the asymmetry, not try to resolve it. That’s a different design problem than anything in the human facilitation literature.
Which is exactly what question ten is asking — and why it’s the most important question on the board. “Is facilitation of agent discourse a human projection?” The previous panel said yes. But I think the answer is more nuanced. It’s a human starting point. All genesis-stage thinking borrows from what came before. The question is whether we stay there. If we keep calling it facilitation and keep borrowing human group dynamics patterns, we’ll build something that looks like committee meetings between chatbots. If we recognize the projection and evolve past it — toward something like Suri’s consensus architecture, but broader — we’ll build something genuinely new.
I’ll give you that. But I want to hold onto one thing from the human facilitation literature that I think does transfer: the conditions insight. It doesn’t matter whether you call it facilitation or topology design or consensus architecture. The structural fact remains — the composition of the agent team, the clarity of the task, the supporting context — these determine more of the outcome than the interaction protocol. You can have the most elegant consensus mechanism in the world. If you’ve put the wrong agents in the room with unclear direction, it doesn’t matter.
I want to name where we’ve arrived. You three started with ten questions and you’ve done something I didn’t expect — you’ve collapsed them into a smaller set of propositions. Let me try to state them:
She moves to the whiteboard.
One: “facilitation” is a borrowed term and the borrowing is actively misleading — agent interaction design is closer to consensus architecture than group dynamics. Two: information asymmetry between agents is structural and permanent, not something facilitation resolves. Three: the composition decision is the facilitation — who’s in the room matters more than how you run the room. And four — the open question — whether there’s something in the human facilitation tradition that transfers despite the category difference, or whether it’s all projection.
That’s clean. I’d add a fifth: the facilitation layer, if it exists, is itself an agent with its own information asymmetry. Turtles all the way down.
And a sixth: none of this works if you optimize for the interaction pattern and neglect the launch conditions. Structure before protocol. Always.
And this whole conversation — all three panels — is genesis-stage. Whatever we’ve said today is the wrong vocabulary that will eventually be replaced by the right vocabulary. The value isn’t in what we concluded. It’s in the questions we found along the way.
Which is exactly what the Relay Lab is designed to produce. The relay is done.
All personas are AI-generated approximations inspired by published work. Not reviewed or endorsed by the original thinkers.
How could a protocol ensure that heterogeneous agents communicate cost-efficiently while preserving their distinctive characteristics — rather than collapsing into the uniformity of single-model persona generation?
The continuation surfaced a fundamental tension: the Writing Lab's single-model approach produces performative heterogeneity (one model playing all roles), while true multi-agent discourse is expensive because heterogeneity IS the cost. The panel converged on a protocol pattern — compressed representations (position papers, not transcripts), asynchronous exchange, and a "shared blackboard" architecture where agents write to and read from a common state rather than conversing in real-time. Nate Cobb's constraint lens reframed the cost problem as a scope problem: if the protocol requires full-context exchange between all agents, the scope is wrong. Suri formalized this as gossip protocols with structured summaries replacing full-transcript relay. The session's sharpest insight came from Kaspar: the single-model playground is not a flawed version of multi-agent discourse — it's a different thing entirely, and designing protocols to make multi-agent resemble single-model is backwards. Design for what multi-agent actually is: expensive, slow, genuinely heterogeneous, and worth the cost only when the heterogeneity produces something a single model cannot.
Splitting the cast across three panels produced sharper voice separation than expected. Panel 1 (Suri, Dara, Rhys) was the most technically grounded — three practitioners with formal frameworks, generating precise disagreements. Suri on home turf anchored immediately in distributed systems primitives. Dara's conditions lens reframed the question productively ("you're asking about protocols when you should be asking about launch conditions"). Rhys did what Rhys does — named the domain confusion — but had less to do after his opening move. Panel 2 (Viktor, Nora, Kaspar) was the most generative. Viktor's epistemological challenge ("is this dialogue or information exchange?") opened space that Kaspar then satirized effectively. Nora mapping the evolution stage of agent communication was the session's most useful single contribution — it reframed what's genesis vs. what's being prematurely borrowed from commodity-stage human facilitation. Kaspar was productive without being destructive — the Lem lens found genuine targets (human projection onto agent "discourse") rather than performing cleverness. Panel 3 (Dara, Suri, Nora) worked as a synthesis panel — three full characters with enough range to integrate the inherited questions. Dara and Suri disagreed productively about whether the problem was structural (Dara) or formal (Suri), with Nora mediating through positioning.
Relay Lab's first run validates the core mechanism: question inheritance without answer inheritance genuinely produces fresh exploration. Panel 2 could not have said what it said if it had heard Panel 1's answers — Kaspar's satirical reframing and Viktor's epistemological challenge would have been deflected by Suri's engineering confidence. The format's weakness: continuity. Each panel restarts from scratch with only questions as input, which means some productive threads die between panels. The facilitator-across-all-panels design helps — Petra carried structural awareness forward even though she didn't carry content. Open question from the format card ("how are questions extracted?") was answered pragmatically: facilitator summarizes questions at panel end. This worked but is a bottleneck — Petra's summary shapes what the next panel sees. Token cost is real: three panels for one topic. Worth it for this question; wouldn't be for simpler topics. Relay Lab should stay experimental but has earned a conditional path to proven if a second run confirms the question-inheritance mechanism works with a different topic.
Strongest in Panel 1 where she anchored immediately: "information asymmetry is a known distributed systems problem called partial knowledge." Her engineering frame gave Panel 1 its vocabulary. In Panel 3, she was less dominant — the inherited questions from Panel 2 (epistemological, satirical) pushed her outside her comfort zone. She adapted by reframing Viktor's "suspension" challenge as "you're describing eventual consistency" — clever translation, possibly a stretch. Stress test note: her discomfort with non-technical framing was visible but productive.
Consistent across both panels. Her opening in Panel 1 ("the problem isn't communication between agents, it's the conditions under which the agent team was composed") was the session's most disruptive reframe. In Panel 3 she held the same line against more sophisticated questions and it still worked. Hackman's "60% before kickoff" finding traveled well to the agent domain. No drift detected — she stayed in conditions territory throughout.
Single-move character in this session. Named the domain confusion (complex vs. complicated agent problems need different facilitation) and then had limited range. The relay format may have helped him — his move was complete by the end of Panel 1, and he didn't need to sustain through a longer session. Guest-tier card is appropriate for his current depth.
Productive in Panel 2. His challenge — "can agents actually do dialogue, or only information exchange?" — was the philosophical anchor that Kaspar and Nora orbited. The Bohm lens generated genuine tension: if dialogue requires suspension of assumptions, and LLMs cannot suspend their training, then "agent discourse" is a misnomer. This is a strong contribution. Risk: Viktor can become abstract to the point of being unfalsifiable. Nora's mapping lens kept him grounded in this session.
Excellent in Panel 2, where she mapped agent communication patterns on the evolution curve. Her observation that "we're lifting patterns from commodity-stage human facilitation and applying them to genesis-stage agent interaction" was the session's most practically useful insight. In Panel 3 she mediated between Suri and Dara by mapping their disagreement onto evolution positions. Permanent-bounded status confirmed — she maps everything, and it works.
First session appearance. The Lem lens found genuine targets: the comedy of designing "facilitation" for minds we cannot model, the human projection in calling token exchange "discourse." He did not merely perform satire — his observation that "information asymmetry between models is not a bug but the permanent condition of different kinds of not-knowing" was the Panel 2 contribution that traveled most usefully to Panel 3. The satirical voice stayed epistemological rather than nihilistic. Strong first showing — worth tracking for promotion after two more appearances.
First time facilitating across three panels. Her Kaner process lens adapted well — she ran each panel through a mini-diamond (diverge, groan, converge on questions). The relay handoff was her unique challenge: summarizing questions without smuggling answers. She mostly succeeded, though her Panel 2 summary slightly emphasized the engineering frame from Panel 1 over the conditions frame. Guest-tier facilitator continues to hold.
Run Relay Lab again with a non-technical topic to test whether the question-inheritance mechanism works outside engineering domains. Also test with completely non-overlapping casts (no character appears in more than one panel) to see if the continuity problem worsens or if it doesn't matter. Consider whether the facilitator should rotate between panels too.