Latent Space

Akshay Nathan on One Shared Harness, Two UX Shells

Akshay Nathan· Core Product Engineering lead at OpenAI at OpenAI
·~71 min·English·Latent Space
AgentsAI CompanyBusiness StrategyAI Infrastructure
TL;DR

OpenAI merged Codex's agent power into ChatGPT Work behind one shared harness after non-developers quietly adopted the coding agent, betting that as AI blurs job roles the product should route users rather than box them in.

01Origin Story

Non-Developers Had a Superpower

The push for ChatGPT Work started when non-developers inside OpenAI quietly adopted Codex and acted proud, like they had been handed a superpower they were not supposed to have.

the thing that really stuck out to me is how proud people were that they were using Codex

— Akshay Nathan, Latent Space
Key Insight
The signal OpenAI read was not a usage chart but a social one: people broadcasting that they were using a coding tool they were not supposed to touch. Pride, not adoption numbers, is what flags a capability that has crossed from niche tool into felt superpower.
02Harness Engineering

One Harness, Two Faces

Codex and ChatGPT Work are not two engines; they run the same underlying harness, and only the UX layer on top differs by mode.

So the harness is the same the harness is shared.

— Akshay Nathan, Latent Space
Key Insight
Sharing the harness is a strategic bet, not just tidy engineering. Because plugins, computer use, and artifacts live in the common core, an improvement there tends to strengthen both experiences at once, while each mode keeps its own UX and sandboxing opinions, which is how developer-grade power can reach knowledge workers without rebuilding it per surface.
03Product Philosophy

Don't Box the User In

Because AI blurs the line between coding, strategy, marketing, and planning, OpenAI keeps the primitives unified and lets the model route users instead of splitting products by identity.

we want to be prescriptive about when to be in either experience, but we don't want to box anyone in.

— Akshay Nathan, Latent Space
Key Insight
The unification is a hedge against the company's own pace of change. If a person's job reshuffles every few months, any boundary drawn around a job title is obsolete the day it ships, so the durable design binds the product to tasks rather than to identities.
04Harness History

Divergence, Convergence, Repeat

Harnesses evolve in cycles: the ChatGPT harness optimized for latency and personality, the Codex harness optimized for a flexible computer environment, and knowledge work is now pulling them back together.

there's sort of like a a constant process of like divergence, convergence, divergence, convergence.

— Akshay Nathan, Latent Space
Key Insight
Framing harness design as an oscillation reframes 'which harness wins' as the wrong question. Today's convergence onto the Codex harness is a phase, and a specialized chat harness tuned for latency and personality will diverge back out the moment a use case demands it again.
05Design Under Complexity

Hide the 32 Knobs

With dozens of model and reasoning combinations available, OpenAI's answer to choice overload is an opinionated default that is best for most people, with the knobs tucked into an advanced view.

we want the default to be the best possible. Like we want to be opinionated about the default

— Akshay Nathan, Latent Space
Key Insight
The single slider is a philosophy rendered as UI. Collapsing model class, reasoning level and sub-agent settings onto one speed-versus-quality axis concedes that most users do not want control, they want a good outcome, and that exposing power should be a tax charged only to the users who opt into it.
06Adoption

Show, Don't Tell

The real bottleneck to adoption is not capability, it is discovery, so the product's job is to show users what is possible in the moment rather than teach them through docs and onboarding.

how can we build the right system such that we can show them instead of telling them what can be done?

— Akshay Nathan, Latent Space
Key Insight
This quietly inverts the DevRel playbook, which happens to be Akshay's former field. When the model can tailor a demonstration to each user's history and exact moment, showing scales in a way human telling never could, which is why he treats the discovery gap, not the capability gap, as the frontier.
07Artifacts

The Site Replaces the Deck

Teams inside OpenAI are dropping slide decks and spreadsheets for generated HTML sites as the canonical shared artifact, because a site is infinitely flexible where fixed tools hit a wall.

But in the case of a site you can do anything you ask for anything and you can get that.

— Akshay Nathan, Latent Space
Key Insight
The shift is really about where the flexibility ceiling sits. PowerPoint and Excel cap out at whatever features they shipped, but an agent generating raw HTML has no feature list to be limited by, so the artifact becomes whatever the task needs instead of whatever the tool allows.
08Measuring Productivity

Motion Is Not Progress

AI makes motion cheap, with more tokens, more pull requests and more dashboards, but progress still demands a deliberate view of what you are actually trying to achieve, and old proxies like lines of code are breaking down.

the trap is like conflating motion and progress. I think motion is much easier now than ever before because of the tooling that we have. But progress requires you to be like very prescriptive and deliberate about like what you're actually trying to achieve.

— Akshay Nathan, Latent Space
Key Insight
The proxy collapse is the deeper point. Story points and commit counts were always stand-ins for 'did we hit the goal,' and AI breaks that correlation by making the proxies trivial to inflate, forcing teams to measure the outcome itself, at-bats with quality, rather than its shadow.