We’ve come to call emerging AI models “frontier AI.”
Yet, the execution of these models within a particular context is like a frontier of its own.
In his essay “On Fairy Stories,” JRR Tolkien described fantasy stories as “sub-creations,” stories that participate in reality—even when they operate within a distinct, fantastical world.
I’d like to call the inner-workings of a model in a particular context a “sub-frontier,” a world marked by the activity of frontier agents participating in a goal mediated by a human author.
You can think of an agent harness, an AI-enabled API layer, etc. as the “sub-frontier.”
Like sub-creations, we interact with sub-frontiers through words and rely on our imagination to apprehend the activity and how it relates to the immediate chapter/task, as well as the end story/goal.
Movies help animate sub-creations, making them tangible to us.
Graphs help animate sub-frontiers, making them more tangible and predictable.
Yet, an avid reader knows that movies cannot do justice to one’s imagination.
Likewise, the best way to understand a sub-frontier is to pin an individual to it who can use words to describe what they imagine is happening.
The best way for an organization to imagine the costs and impacts of AI integration—a common problem—is to surface the summaries of individuals who imaginatively author the sub-frontiers.
This makes the process incredibly simple: empower the people closest to a sub-frontier to act decisively and trust their summaries as the metrics that inform reporting by higher-ups.
The pre-AI machinery of structured reporting mechanisms lose value in a post-AI landscape.
Concretely, this looks like having a single individual responsible for every agent, and giving every agent a clear goal.
Each slice of an organization would then have a team of people, if needing more than one individual, that corresponds to a team of agents (and their sub-frontiers).
This collection of sub-frontiers (i.e., team of people/agents) forms a colony representing the activity of an organizational slice.
The costs and value of a sub-frontier map to a person’s account.
The costs and value of the colony map to the collective accounts of the team members.
The costs and value of the company map to the costs and values of the colonies.
The anti-pattern should also be clear:
Agents cannot be pinned to multiple people.
Colonies cannot be pinned to multiple teams.
Costs and value are not determined through reporting machinery, but rather through the plain summaries of people responsible for an agent.
There cannot be multiple agent technologies/platforms at play.
Software systems should not be centralized if they create a bottleneck—this should be re-evaluated according to the new organizational structure.
The goal in this transformation is reached if organizations can grasp costs and value, with a direct path to who is responsible for what.
If the restructuring doesn’t attain that goal, then AI cannot reach its payoff.
This far in, it is worth a shot.