Multi-agent organization design

Independent evaluation must be a separate role with the authority to kick work back, because an agent grading its own output is not a check. And translation into human-legible language is itself a role: inter-agent communication is not human-readable by default, so a summary is a function someone must own rather than a byproduct of the work. Both show up in the same production framework at a regulated bank, arrived at without reference to anyone else’s methodology.

One camp scales agent count and gives the population explicit organizational machinery — group meetings, divisions, an agent school, self-organizing encounter spaces — and reports results that small teams did not reach, with humans participating as ordinary members rather than as operators. The other camp reports that swarms recreate microservices failure modes: compounding cost, compounding confusion, a testing burden at every handoff, fine-grained personas that burn tokens negotiating instead of working. That camp caps the roster at about four well-defined agents and treats the context handoff as the real design problem.

Declare the topology as data — each node stating what it does, what it needs, and who it may call — so the glue code disappears and the structure becomes testable and machine-authorable. Route work down tiers until a cheap model handles each leaf reliably. The claim behind it is that after years of multi-agent work, the agents-calling-agents glue was “the noise”, and lifting it into configuration lets test harnesses be generated rather than handwritten — and lets multi-agent systems author other multi-agent systems as disposable, perishable thoughts.

The enterprise pattern is that roles multiply because someone has to be answerable for each one, not because the technology suggests it — tenants running three or four hundred discrete “AI employees”, with “AI employee manager” now a real job title. The adoption path is task by task rather than role by role, on the finding that the blocker is never the agent; it is the humans collaborating with it. And the usage data cuts against the demo: about half of builds are conversational, but over ninety per cent of invocations are programmatic. Conversational agents get built; triggered ones get used.

Also: BMAD; Spec Kit; Kiro; Neuro SAN; the Berkeley Function Calling Leaderboard; ARC-AGI.

Assay’s four-desk pipeline maps almost one-to-one onto the bank’s four agent classes: understanding is intake-desk, planning is brief authoring plus worker-desk, independent evaluation with kick-back authority is PR review plus verify, and the explainer is the pull-request and issue prose layer. That the same shape was arrived at independently inside a regulated bank is the strongest external evidence in this scan that it is not idiosyncrasy; the consultancy’s plan → component → developer chain is the same convergence from enterprise legacy work. Critically, the role separation is enforced by identity, not convention — each desk posts as its own GitHub App, so self-review is structurally impossible because the reviewer is a different principal.

Five desk roles, not fifty; one brief, one branch, one pull request. The handoff that the swarm-skeptics call the real risk is managed with durable claim records on the forge, narrow verbs and a single generated board. The four-agent cap and the handoff finding are convergent external support that those controls are load-bearing rather than bureaucratic.

First, the explainer as a separate class says human-facing summaries are a distinct function rather than a byproduct of worker output; Assay treats pull-request bodies and issue prose that way in practice but has never named it as a role. Second, the entropy finding independently justifies separate worker and reviewer identities for a reason Assay had not claimed: model diversity supplies variance that one model refining its own output cannot. Assay routes by role, not by task difficulty, and sizes briefs so an ordinary worker model succeeds rather than so a genius model can hero through — a deliberate difference from the routing camp. Third, declared topology is the closest thing to a critique that lands: briefs and skills are Assay’s declared structure, but nothing declares a per-node upward contract — what I do, what I need, who I may call — which is what skill metadata could carry so a linter could verify call-graph legitimacy. Designed here, unbuilt.

The hallucination-layer rule — source it, explain why you cannot, or drop it — is Assay’s evidence-citation discipline arriving from an independent production context. And “the blocker is the humans, so adopt task by task” is the right frame for rolling the methodology out to skeptical repositories: desk by desk, not all at once.

• Whether the swarm results reproduce outside science domains with cheap external validators, or stay confined to fields where a wet lab settles the answer.
• Whether anyone measures the bounded-roster claim directly: cost and error rate at 4 agents versus 40 on the same task.
• Whether declared-topology configs get adopted beyond their authors, which is what would make a machine-verifiable call-graph contract worth building here.