Latent Space

Jay Parikh on rebuilding every layer of GitHub for agents

Jay Parikh· EVP of CoreAI at Microsoft at Microsoft
·~23 min·English·Latent Space
AgentsAI InfrastructureBusiness StrategyOpen Source
TL;DR

Microsoft's CoreAI lead walks through what it takes to run GitHub when agents already open 40% of pull requests: rebuild every layer, make the old workflow optional, route across models, keep the quality bar from slipping, and tie token spend to outcomes.

01The Shift

You Don't Have to Build Alone

<strong>Agents doing the grunt work widened GitHub's audience past skilled engineers</strong> — and the human count still grew to 225 million, one new developer every second.

you don't have to be a, you know, a skilled software engineer to use GitHub, right? So now with coding agents can do a lot of things for you and then you can share an idea.

— Jay Parikh, Latent Space
Key Insight
The headline stat is agents, but the real story is distribution: when agents handle the engineering, GitHub stops being a tool for a profession and becomes a place anyone can ship from. That reframes the 225M number as a floor, not a ceiling.
02Workflow

The End of the Pull Request?

<strong>GitHub's answer to "agents killed the PR" is to make the workflow optional, not mandatory</strong> — the platform ships concepts you assemble however you work.

I don't want to deal with issues or PRs anything right so again the platform should allow and have the affordances for these things so we'll keep adapting

— Jay Parikh, Latent Space
Key Insight
Parikh won't declare the PR dead, but he won't defend it either. Treating the pull request as a configurable affordance rather than a law is how a platform at 225M users absorbs a paradigm shift without forcing every team through it at once.
03The Harness

Fusion Takes the "Which Model?" Decision Away

<strong>Copilot's harness orchestrates the tools; Fusion routes the task to the right model by price and performance</strong> — so the engineer stops shopping for models.

Hey, we can just like optimize that for you. We can get you the best quality really fast and much more cheaply.

— Jay Parikh, Latent Space
Key Insight
The implication reaches past convenience: once Fusion owns the price-performance decision, the specific frontier model matters less to the user than the harness that chooses it — the value concentrates in the router, not in any one model.
04Product Strategy

One Harness, Every Surface

<strong>Phone, desktop, VS Code and github.com now share one harness, so context compounds in a single place</strong> — the thing stitched-together third-party tools can't do.

if you're using one tool from one other provider here and then you flip to another one, they don't compound their intelligence together

— Jay Parikh, Latent Space
Key Insight
Stated as a feature, this is also the lock-in argument. Parikh only claims portability and accumulated preferences — but the implication is that the longer context compounds in one place, the more a switch would make you give up.
05Open Source

The 1980 Problem

<strong>Dropping 2,000 dependencies to 20 doesn't erase the maintenance — you inherit the other 1,980 yourself</strong>, which is why Parikh argues for securing open source rather than abandoning it.

let's solve the problem, which is like let's make the stuff in open source secure versus everybody going into like some walled garden approach.

— Jay Parikh, Latent Space
Key Insight
The "code is cheap, so rewrite everything" pitch sounds like a security win until you count the maintenance. Parikh reframes the choice: cheap generation doesn't retire the dependency problem, it just decides whether the whole industry or every team alone carries it.
06Security

The Quality Bar Can't Slip

<strong>Generating far more code only works if the gates get stronger</strong> — code review, AI deep scanning, and secret-blocking stand between agent output and the repo.

we're generating so much more code. We want to make sure the quality bar doesn't slip. things like code review, copilot code review help you maintain that quality bar.

— Jay Parikh, Latent Space
Key Insight
When volume goes up and the author is often an agent, human review can't scale with it. Parikh's security story is really a throughput story: the only way to keep quality flat while output explodes is to automate the gates that used to be a person's job.
07FinOps

Show Me the Tokens Are Spent Well

<strong>GitHub is building spend governance the model labs won't — because labs want you to spend more.</strong> The ask from enterprises is limits, visibility, and spend mapped to outcomes.

especially the fortune 500 they want governance they want to be able to put in limits you know whatever so we have to provide the controls the governance there

— Jay Parikh, Latent Space
Key Insight
Parikh is drawing a line between a platform and a model vendor. A vendor's incentive is more tokens; a platform's is spend the buyer can defend. Read that way, the cost controls the labs skip are exactly what an enterprise needs to tie token usage to outcomes it already tracks.
082027 Prediction

Designing for an Unknown Number of Agents

<strong>Parikh won't say whether next year means dozens, hundreds, or thousands of agents — so GitHub is building composable primitives that hold at any of those scales.</strong>

Are we managing thousands of agents versus, you know, dozens or hundreds? I don't know.

— Jay Parikh, Latent Space
Key Insight
The honesty is the signal. Parikh won't name the number, but he's confident enough in the direction to design for it — composable primitives are a hedge against not knowing whether you're orchestrating tens or thousands of agents a year from now.