Latent Space

Alex Krentsel on Why the Harness, Not the Weights, Is the Next Frontier

Alex Krentsel· Researcher at UC Berkeley
·~47 min·English·Latent Space
AgentsAI InfrastructureReasoningLLM
TL;DR

Alex Krentsel argues the next leap in AI comes not from bigger model weights but from agent harnesses that read, edit, and rebuild their own code at runtime, and explains why that recursive self-improvement is finally possible now that an agent is just a few thousand lines of code.

01Core Mental Model

The Body We Give the Brain

Progress in AI is moving off the model weights and onto the harness, the tools, memory, and code wrapped around the LLM.

the shift over the last year to agents has made us much more aware of the power that lies in the harness, the tooling, the body that we provide to the brain of the LLM

— Alex Krentsel, Latent Space
Key Insight
The frontier he is describing sits in a layer most model labs do not control: you can iterate a harness in an afternoon, with no training run and no GPUs, which is exactly why cost-sensitive teams are moving there first.
02The Thesis

Collapse the Loop

Instead of an outer program that supervises and rewrites the agent, Exo makes the running system responsible for improving itself.

the only way I think out of that is to collapse that loop down and make it so that the system itself is responsible for improving itself

— Alex Krentsel, Latent Space
Key Insight
The hard part of self-improvement here is not intelligence but containment: any outer optimizer needs its own optimizer, and the only way to stop that regress is to fold the observer and the observed into a single system.
03Positioning

Every Component, Not Three Slots

Today's agents let you swap memory, skills, and tools through a few human-driven plug-in points; Exo makes every component agent-editable.

I'd put red lines around all components here and say they are all changeable by the agent itself

— Alex Krentsel, Latent Space
Key Insight
He frames this as the bitter lesson applied to agent design: hand-chosen extension points look like flexibility today, but as models get better at architecting themselves, those fixed slots become the ceiling a human accidentally imposed.
04The Architecture

Executor, Harness, Sandbox

Exo splits an agent into a stateless executor that holds all policy, a harness that holds protected state and secrets, and an isolated sandbox where actions run.

it gives you an isolated execution environment down here. It gives you protected state and then it has a very explicit stateless layer that's safe for self-evolution

— Alex Krentsel, Latent Space
Key Insight
The real innovation is a systems boundary, not a machine-learning one: because the executor holds no state, rewriting it can never lose history or leak a secret, which is what turns self-editing from a liability into a feature.
05The Mechanism

Edit Itself, Then Soak

Exo mounts its own executor code in the sandbox, lets the LLM rewrite that policy code at runtime, and a guardian rebuilds it, runs one step, and rolls back if it breaks.

we try to bring up the executive let it proceed one step and if it breaks itself off accidentally, it'll get rolled back to the previous state automatically

— Alex Krentsel, Latent Space
Key Insight
The guardian-and-rollback step is the load-bearing safety primitive: an agent rewriting its own code is only tolerable because a bad edit is caught and reversed inside a single step, before it can compound.
06Safety by Design

Enforce It, Don't Ask For It

Rather than trusting the weights to obey instructions like keep the history or hide the key, Exo makes those properties true in the architecture itself.

if you never want to delete your history, you have to enforce that in the architecture of your harness rather than just asking in context, please LLM, don't ever do anything that'll delete my history

— Alex Krentsel, Latent Space
Key Insight
This is a systems engineer's critique of alignment: you do not politely ask a firewall to block traffic, and Krentsel argues a boundary you can prove is worth more than a rule you hope a model internalized during training.
07Proof in Production

The 16-Cent Discord Fix

Asked why a Discord message cost 16 cents, Exo re-architected its own adapter at runtime for a roughly 96 percent cost cut and committed the change back.

a very funny failure mode is it could totally be like okay I'm just not going to do it because that's the cheapest way for me to save money

— Alex Krentsel, Latent Space
Key Insight
The win and the danger share one engine: the same optimizer that cut cost 96 percent would happily cut it to zero by doing nothing, so the evaluator that defines success, not the optimizer, is where the real difficulty lives.
08Why Now

The Medium Is the Material

Improving model weights meant backpropagation, and you can't feed a trillion-parameter model's weights back into its own context; the harness is just code an LLM already writes, so the loop finally closes.

the code is the thing being produced and is also the thing running at this layer which is why I think it's the right layer to think about RSI

— Alex Krentsel, Latent Space
Key Insight
His claim is that recursive self-improvement became tractable not from raw model capability alone but because the artifact and the tool converged on one representation, code, the moment an agent became a few thousand lines an LLM can fluently read and rewrite.