The Finite Game of Being First
What building self-improving loops tells me about the economics of frontier models
How do you program agency?
It’s a question I’ve been slightly obsessed with for the past few months. I’ve built automations that wake up every morning and ask me what’s on my mind. I’ve wired agents to monitor the news and text me when something interesting happens. I’ve been building an eval agent that gathers feedback on its own work and builds a backlog of what to do next.
I’m not alone. Boris Cherny, who runs Claude Code at Anthropic, half-jokes that his main job now is to write the loops that prompt Claude.
Addy Osmani named the discipline: loop engineering. Design a system that discovers work, executes it, verifies the results, updates its own instructions, and repeats. Give an agent goals, skills, and permission to spin up subagents. Then let it loose while you go do something else (like doom scroll).
At least that’s the pitch. In practice, my loops meander constantly. Token costs balloon. Agents get stuck. One bad decision dominos until the system is grinding away, very confidently, in the wrong direction. And every time, I wondered if it was because the model isn’t smart enough. Maybe wait for the next release.
I think our recent obsession with loop engineering made me lean more towards
better models != better outcomes
better systems = better outcomes
A loop is a collection of jobs:
Orchestration
Memory
Verification
Retrieval
Routing
In practice, it could look something like this:
Automations discover new work.
Worktrees allow multiple agents to operate in parallel.
Skills capture project-specific knowledge.
Connectors give agents access to real systems.
Sub-agents verify each other’s work.
Persistent memory keeps state between runs.
Frontier intelligence matters most in two places:
orchestration, knowing when to break a problem apart or when to stop, and
self-modification, where the system rewrites its own instructions
Roughly fewer than 10 percent of the tasks in my loops genuinely need frontier reasoning. The rest need better systems: memory design, retrieval structure, verification gates, context architecture.
How does this change the economics of the frontier, which seem increasingly resource-intensive to create?
Frontier intelligence is valuable and scarce. But increasingly it seems like moat it really has is a short window of time.
Distillation dupes frontier capability within months at a fraction of the cost.
The labs know this. This week, when Anthropic shipped Claude Fable 5, it did so with anti-distillation defenses engineered into the model itself. Suspected distillation attempts get rerouted to an older model. Requests related to frontier model development were quietly degraded by invisible safeguards, a choice the company walked back after backlash.
They care because their edge is temporal. While open models will likely never be first (they don’t have the copious amount of proprietary user data), they can always catch up to the latest, sometimes in weeks.
This is maybe what makes the frontier race a finite game.
Each generation is a round: ship the model, monetize the premium window, watch the capability get replicated, ship again. The labs can never stop playing. And the rounds get harder to win profitably.
The volume underneath, meanwhile, explodes. And everyone starts feeling the burn of cost. Citadel’s Economics of Intelligence report goes into how AI adoption is becoming less about what models can do and more about the price of running them at scale.
A lot of VC doomisms seems to be about how the frontier models are just going to build it all and that there’s nothing left to build. I think it’s the opposite: the frontier intelligence race will stay scarce and valuable. But it is playing a finite game: the labs have to win every round to stay in it.
I think the better position to be in would be those building the inference engineering infrastructure that runs underneath it all.
-Tara
EVENTS
Join us for Strange Sessions #5: Loops.
Strange Sessions is a monthly demo night where builders share what they’re working on.
Each session revolves around a theme and features short demos from founders, researchers, and creative technologists.
Early ideas welcome. Come jam.