Dust Multi-Agent Orchestration vs Sup AI

“The primitive here is a directed task graph where agents can spawn sub-agents with scoped permissions — that's a real primitive, not a marketing word. The DX bet is that you configure agent topology in a UI rather than in code, which is the right call for enterprise buyers who don't want to version-control YAML agent graphs. My concern is the moment of truth: connecting your first data source and actually watching agents delegate requires significant setup around connectors and permissions, so the first-10-minutes test is rocky. Still, this isn't a three-API-call Lambda wrapper — the audit trail and scoped delegation are non-trivial to build correctly, and Dust appears to have built them correctly.”

“No API, no self-hosting option, and the ensemble approach means your per-query cost is 3-5x a single model call. The benchmark numbers are compelling but I cannot integrate this into a product. Ship an API and I will reconsider.”

“Direct competitors are Salesforce Agentforce, Microsoft Copilot Studio, and ServiceNow's AI layer — all of which have distribution advantages Dust will never replicate. The specific scenario where this breaks is any enterprise with a non-standard data stack: if your knowledge lives in a homegrown CRM or an obscure ERP, Dust's connector set will leave you writing custom glue code that defeats the point. What kills this in 12 months isn't a competitor — it's that Anthropic and OpenAI both ship native multi-agent orchestration APIs that remove Dust's orchestration layer as a distinct value prop, leaving only the compliance UI as a moat, which is thin. To stay alive, Dust needs to own the compliance and audit workflow so deeply that even when orchestration is commoditized, enterprises can't migrate without losing institutional governance history.”

“The benchmark result is legitimately impressive and the methodology is transparent. My concern is latency — querying multiple models and aggregating adds significant time. For research and high-stakes questions it is worth the wait. For everyday chat it is overkill.”

“The buyer here is the Chief of Staff or VP of Operations at a 500-1000 person company, pulling from a digital transformation or IT budget — that's a real check-writer with a defined problem. The pricing architecture is opaque (contact sales for anything serious), which means every deal is a negotiation and CAC balloons, but enterprise SaaS lives or dies on ACV so this is forgivable if they close at $50k+. The moat is the audit log and permission graph embedded in workflows — switching costs come from compliance teams relying on Dust's logs for actual regulatory reporting, not just convenience. The risk is that the underlying model providers ship governance primitives natively, collapsing Dust's differentiation to UI, which is not a durable position.”

“The thesis Dust is betting on: by 2028, enterprises will run hundreds of specialized AI agents simultaneously, and the coordination layer between them — not the agents themselves — becomes the strategic chokepoint. That's a falsifiable claim, and the dependency is that agent task complexity scales faster than any single model's ability to handle it in one context window, which is plausible given how context window gains have plateaued relative to task complexity growth. The second-order effect that matters isn't productivity — it's that the audit log becomes a new kind of organizational memory, and whoever owns that graph owns the institutional knowledge layer. Dust is riding the enterprise compliance-meets-AI trend, and they're early enough that the design space isn't locked — but the window closes fast once platform players treat orchestration as a checkbox feature.”

“Confidence-weighted ensembling is the quiet breakthrough everyone is sleeping on. Individual models plateau — but smart aggregation keeps pushing the frontier. Sup AI scoring 52% on Humanity's Last Exam when no single model breaks 40% proves the thesis.”

“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”