Agent networks: if capability lives in the orchestration, so does accountability
A new essay from the DeepMind Institute by Benjamin Bratton and Blaise Agüera y Arcas (The Decoder also credits James Manyika) argues that AGI may not arrive as a single model at all. Their observation is that the best AI agents today are frameworks that divide work among several models, and that general capability may emerge from societies of agents, tools, institutions and people working together. In their words, the central problem becomes how to "orchestrate, govern, and live within a complex network of AI agents, people, and the systems that connect them.
Two lines in the essay matter most for anyone who has to answer for these systems. The first is that an agent "appears singular, but is a multitude", a temporary assembly of models, memories and tools that keeps re-forming around its context. The second is that the collective intelligence of these networks will sit in "the embedded rules, procedures, precedents, and feedback mechanisms of the scaffold" rather than in any one agent, with orchestration harnesses that can often outperform single models that are nominally smarter.
That is the authors' argument. What follows is mine.
If capability lives in the orchestration, accountability has to live there too. Most governance conversations still ask whether a given model is safe, and that question gets less useful when the behaviour that matters comes from how many models, tools and permissions were wired together. A regulator or a board cannot inspect an assembly that reforms itself, but they can ask who designed the orchestration layer, what it is permitted to do and who answers when it does something nobody intended.
The essay also says performance is not tied directly to who or what occupies a role, whether a human, an AI or a composite group. I agree for output quality and disagree for accountability. A role can be filled by anyone, but someone named has to own the system that fills it.
Two questions for Australian leaders. For the agent-based systems you run today, can you name the single executive who owns the whole system, not just the vendor model inside it? And if that system changed its own configuration tomorrow, who would find out first?
#AustralianBusiness #Governance #Boardroom
Source: Bratton and Agüera y Arcas, "Artificial symbiotic intelligence: Agents, AGI and the orchestration of many minds", DeepMind Institute, https://lnkd.in/gBXgPprM