30% of Code Has Been Taken Over by AI — What Can Engineers Still Do?
On the evening of April 29 (U.S. time), during Meta’s LlamaCon 2025, Mark Zuckerberg (CEO of Meta) and Satya Nadella (CEO of Microsoft) appeared on stage together for the first time — and they made the same bold declaration:
“Engineers are no longer people who write code.
They are commanders of AI teams.”
GitHub Copilot has evolved into a “prototype agent,” capable of generating pull requests and managing workflows independently.
Inside Meta, each developer is now paired with several small AI “apprentices” working under their direction.
Their shared conclusion was simple:
“We’re not eliminating engineers — but we are completely recompiling their roles.”
This article breaks down that conversation through key questions:
- After AI writes 30% of the code, what’s left for engineers to do?
- What does it actually mean to “lead an AI squad”?
Part 1｜AI Hasn’t Taken Jobs — It’s Changing What “Work” Means
“Inside our company, up to 30% of new code is now written by AI.”
— Satya Nadella, LlamaCon 2025
Nadella explained that Copilot now handles prototype tasks end-to-end — generating PRs, orchestrating workflows.
This is not assistance; it is takeover.
But the deeper signal is not the percentage.
It is that AI is redefining what engineering work actually consists of.
The Deconstruction of the Engineer’s Actions
Zuckerberg shared how Meta structures development:
Developers break tasks into pieces and delegate them to smaller models — for coding, debugging, testing, and committing.
One human engineer steps in only at the end for refinement and review.
AI handles about 80% of the workload.
He called this “the decoupling of engineering actions.”
The repetitive aspects of engineering are being stripped away.
Engineers are shifting from end-to-end builders to task orchestrators.
As Zuckerberg put it:
“AI will do what you tell it to do.
The question is how clearly you tell it what to do.”
AI Doesn’t Just Execute — It Redefines Execution
Nadella added that the real transformation is when AI creates new workflows, instead of accelerating old ones.
For example, preparing for a client meeting used to involve searching, writing, and emailing.
Now Copilot automatically integrates CRM, updates, and communications into real-time proposals.
The workflow itself disappears.
From “Using AI” to “AI Working — Humans Judging”
Both CEOs agreed that developers must move upward into structural design and quality governance.
Zuckerberg predicted:
“By next year, half of Meta’s development will be led by AI.”
And Nadella asked:
“For every line of code — does a human really need to do this?”
Part 2｜Engineers as AI Orchestrators, Not Coders
“Future engineers will be like technical directors — commanding a squad of AI models from design to deployment.”
— Mark Zuckerberg
This shift creates a new archetype:
The Technical Commander
Historically, engineers were valuable because they could do everything themselves.
Now they are valuable because they can coordinate many models to do everything together.
From Hands-On Skills to Orchestration Skills
Meta senior engineers no longer write most logic themselves.
They use platforms to assign subtasks to Llama sub-models, and let Copilot test and deploy.
As Zuckerberg described:
“It’s teamwork — except the team is entirely AI.”
Nadella explained Microsoft’s “Agent Development Stack” — where development depends less on languages and more on which intelligent resources you can command.
AI Tools Become Operational Units
To lead an AI “team,” engineers must master:
- Intent Modeling — turning goals into model-ready tasks
- Model Orchestration — choosing which models do which jobs
- Task Supervision — verifying results and managing errors
Zuckerberg compared it to film production:
- models = actors
- Copilot = assistant director
- APIs = cameras
- deployment = premiere
The identity shift is the hardest:
“Engineers aren’t writing code. They’re writing structure.”
Part 3｜The Most Powerful Model Is the One That Collaborates
“Orchestration” appeared repeatedly throughout the dialogue.
The first generation used one model per task.
Now we are entering the era of multi-model collaboration.
Zuckerberg said developers no longer need one supermodel; they need specialized models that talk to each other.
From Model-Centric to Orchestration-Centric
Gen-1: one large model
Gen-2: many smaller models working together
Modern AI applications look like:
- a chat interface in front
- a swarm of models coordinating in the back
Microsoft is transforming Copilot from a generator into an orchestrator:
- one model interprets intent
- one retrieves data
- one writes code
- one verifies and deploys
Meta’s “Distillation Factory” builds many small, specialized models — an “AI alliance.”
MCP, A2, LoRA — The New Grammar
Nadella introduced two protocols:
- MCP — coordinates task distribution
- A2 — governs communication between models
These will become foundational like HTTP.
Developers will not train one huge model — they will compose systems of models.
Part 4｜Open Source Isn’t About Free — It’s About Control
Zuckerberg emphasized that Llama’s value is in customization.
Meta’s “Distillation Factory” takes a large multimodal model and creates specialized smaller versions for different uses.
Azure is building this into a platform capability.
Open source enables distribution, not just sharing.
The era of massive closed models is plateauing; modular AI is rising.
Part 5｜AI Is Rewriting Organizational Architecture
“We used to organize around people. Now we’re reorganizing around model capabilities.”
Traditional hierarchical structures solved:
- information gaps
- inconsistent decisions
- manual coordination
AI dissolves these by aggregating information, harmonizing decisions, and automating execution.
Zuckerberg calls this “the atomization of organizations.”
Inside Meta, agent teams already replace cross-department meetings.
Managers shift from people allocation to agent architecture.
“Humans become model users.
Models become organizational participants.”
Part 6｜AI Is No Longer a Tool — It’s Your Operating System
Zuckerberg concluded:
“AI isn’t optional. It’s infrastructure.”
Nadella added:
“It’s a new factor of production.”
AI now sits under every workflow and every application.
Applications become:
prompt + orchestration + model structure
Documents become:
conversation + data + interaction logic
Engineers move from writing functions to designing intent structures.
The Real Divide
It is not whether you use AI, but whether you can organize it.
New roles emerge:
- Technical Commander
- Model Orchestrator
- Distillation Engineer
- Agent Architect
Not how fast you code, but how precisely you orchestrate.
The Productivity Paradox
AI won’t raise productivity until organizations rewrite their processes around it — just like factories had to redesign their layouts to benefit from electricity.
Humans must shift from executors to system designers.
A New Beginning
When asked what gives him optimism, Nadella quoted Bob Dylan:
“You’re either busy being born, or busy dying.”
Then he added:
“The smarter choice is to stay busy rebuilding.”
The race is not about model size, but about who can tame AI and embed it structurally.
Engineers evolve into model conductors.
Managers into process architects.
Companies into AI-native organizations.
Only then will “30% written by AI” mark not replacement — but evolution:
Humans no longer as executors,
but as designers and directors of intelligent systems.