A Meta Senior Engineer's Deep Dive into AI-Assisted Programming: Is Hand-Written Code Dying?

A senior software engineer at Meta (L7 level) compiled and systematically addressed over 300 deduplicated community questions following an AMA session, sharing his views on the future of software engineering, agentic programming workflows, and career development in the age of AI. These frontline insights from a top-tier tech company are far more valuable than abstract industry predictions.

Are Software Engineering Jobs Safe?

When confronted with fears about AI replacing programmers, this engineer offered a cautiously optimistic answer. He invoked Jevons Paradox to explain his reasoning: as AI coding tools become cheaper and usage skyrockets, the total amount of code shipped will only grow — which means more engineers will be needed to review, merge, and maintain it.

That said, he acknowledged the other side of reality — the barrier to entry for junior engineers has genuinely risen. "If you're already in the industry, I'd encourage you to stay, keep learning, and embrace these tools. But if you're just starting out, I can't 100% recommend it, because getting in is genuinely hard."

His assessment of current model capabilities is refreshingly balanced: "The latest models produce great output, but also terrible output. Sometimes an LLM is a genius; other times it trips over the smallest thing." This coexistence of brilliance and incompetence is precisely why engineers won't become irrelevant anytime soon.

Jevons Paradox originates from 19th-century British economist William Stanley Jevons, who observed in 1865 that improvements in steam engine efficiency did not reduce coal consumption in Britain — consumption actually surged, because lower operating costs unlocked far broader applications, expanding total demand. This pattern recurs throughout tech history: cheaper storage didn't reduce data, faster networks didn't reduce traffic, and more efficient compilers didn't put programmers out of work. Applied to AI coding tools, the logic suggests: when AI drives down the marginal cost of writing code, companies and individuals will build more software, generating larger codebases — and in turn, more demand for review, testing, maintenance, and refactoring. This is a relatively optimistic take, though worth noting that Jevons Paradox describes aggregate demand expansion, not individual job security. The gains may concentrate among a smaller group of engineers who can effectively leverage AI tools, rather than spreading evenly across the field.

What an Agentic Programming Workflow Actually Looks Like

On day-to-day workflow, this engineer revealed that he primarily uses Codex (especially the Codex app) for most of his work, while keeping Claude Code for complex tasks that require deep processing. His strategy: "Stick with the tools you know well, but try new things when a major paradigm shift happens — like a new model release."

He emphasized that learning by just "doing work" has a ceiling. "I read the docs — I'll go through the entire OpenAI or Claude Code documentation, then try every feature and every shortcut they mention." He believes that deeply exploring a tool's hidden capabilities often leads to unexpected changes in how you work.

How He Reviews AI-Generated Code

Code review is a topic on which he has repeatedly changed his mind. He thinks of a codebase like a tree:

Trunk code: Highly interdependent code that touches core hot paths — a bug here causes major problems. For these diffs, he raises the bar for both his own review and the AI agent's scrutiny.
Leaf node code: For example, newly created components not yet integrated — low risk, reviewed more loosely.

He repeatedly stressed that safety is the primary anchor during review: "I look at gating and security first, then quality and cleanliness." Practical steps include keeping diffs small and atomic (for easy rollback and testing) and requiring agents to provide "confidence proofs" — screenshots, videos, runtime data, unit tests.

One line of his cuts to the heart of accountability: "In a postmortem, you can't say 'the AI wrote it, it's not my fault.' At the end of the day, you're responsible for what you ship."

reviewing AI code workflow

Agentic programming workflows treat AI as an autonomous agent executing tasks, rather than a simple code-completion assistant. Unlike the traditional "ask-and-paste" interaction model, an agent can receive a goal, autonomously plan steps, invoke tools (file I/O, terminal commands, browser actions), and iteratively verify results until the task is complete. Key concepts from the article: MCP (Model Context Protocol) is an open protocol proposed by Anthropic that enables standardized communication between models and external data sources and tools. Sub-agents and agent teams refer to architectures that break complex tasks into specialized sub-agents working in parallel — analogous to microservices in software engineering. Memory mechanisms determine how an agent stores and retrieves context across multi-step tasks. Understanding these underlying mechanisms — rather than just knowing how to use a tool's interface — is exactly the difference between deep mastery and surface-level familiarity that the author emphasizes.

Hand-Written Code Is Dying

When asked "how much code do you write by hand?", his answer was blunt: "Not a single line."

He admitted this transition came with a sense of loss — as someone who came from a non-CS background, spending five years coding as a hobby before switching careers, he went through a "grieving" phase for hand-written code. But he sees it as an unavoidable cost of this transformation: "Anyone telling you hand-written code isn't going away mostly just hopes it won't."

He also addressed the debate over whether "coding is solved": "With enough context and enough validation, the coding part is solved — AI will write the code for you. But that doesn't mean all real engineering work is done."

Staying Sharp After You Stop Writing Code

For the frequently asked question of "how do you keep your programming instincts without writing code?", his approach is a flywheel of continuous learning + teaching output.

He shared his experience learning model scaling as a non-ML engineer: he worked through JaxML tutorials section by section, using Codex and the Claude Chrome extension to ask questions line by line. "Reading one page could take hours, but the depth of understanding afterward was far beyond skimming."

He's been making YouTube videos for eight years, and his core motivation is: "If I want to go deep on a topic, I teach it." "Once you try to explain something you think you understand, you immediately see the gaps in your knowledge."

staying sharp with AI

Competitive Advantages and Job-Hunting Advice in the AI Era

On the question of "if AI lets everyone build, what's the real competitive moat?", his answer centers on domain knowledge and data moats: the social graph of a platform like Facebook, or Tesla's autonomous driving mileage data — these are barriers AI can't easily flatten.

But his more grounded insight is that AI tool proficiency itself is a competitive advantage. "You'd be surprised how few people have really gone deep on these AI coding tools. Most regular people are still at ChatGPT and Gemini — they've never heard of Codex."

For newcomers to the industry, he lays out two tracks:

Master one agentic coding tool: Understand how memory works, what skills are, what MCP is, the difference between sub-agents and agent teams — even being able to "feel" whether output came from Opus or another model.
Build solid fundamentals: Data structures, algorithms, design patterns, Clean Code principles. Because "agentic programming is fundamentally an abstraction, and abstractions are inherently leaky" — when an agent gets stuck, only strong fundamentals let you step in and debug.

job ready skills

A Real-World Path Into Big Tech

He shared his two attempts at joining Meta: the first time he "got absolutely destroyed" — failed every interview. The second time, during paternity leave, he studied four hours a day for a month and a half straight, ultimately landing offers from both Meta and Netflix.

He believes interviewing is in some ways "a solved problem" — the hard part is sustaining the grind until you can actually pass. "NeetCode is great. If you nail 80% of it, you can get into a top company. The key is that when you read a problem, you instinctively recognize whether it's a sliding window, tree, linked list, or graph problem."

For those without a prestigious school background, his path was to build a few years of experience at a regular company first — he was three years into Capital One before Meta and Netflix started reaching out. He also strongly advocates building in public and content creation as a networking lever: "Social media is a level playing field. If you produce valuable content, people will notice you."

build in public

Codex App vs. Claude Code: Why He Switched

On tooling, the engineer revealed an interesting shift — two months ago he was convinced he'd never leave the CLI; now he's almost entirely on the Codex app.

His reasoning: all major companies are targeting the mass market, and app-based interfaces (non-terminal) enable rich rendering, voice control, drag-and-drop task queue editing, and other capabilities that CLI simply can't match. Long-term feature investment will inevitably tilt toward the app. "It's hard to imagine a pure TUI experience ultimately winning."

On model quality, he acknowledged that pricing and rate limits are becoming key factors — which is why he shifted from Claude's Fable to Codex. "OpenAI's model output is very comparable, and there's less fluff. As my friend Peter puts it, it 'bites like a dog' — it keeps going until it gives you an answer."

Codex is OpenAI's code generation model and product line, originally offered as an API and recently released as a standalone app. Claude Code is Anthropic's command-line AI coding tool (CLI) for developers, known for deep codebase comprehension and long-context processing — particularly strong at complex refactoring tasks spanning multiple files and modules. The two represent the two major form factors for agentic coding tools today: CLI tools integrate seamlessly with traditional developer workflows, terminals, scripts, and CI/CD pipelines; app-based tools offer richer visual interfaces, task queue management, and multimodal interaction with a lower barrier to entry. "Fable" mentioned by the author is a mode within Claude Code designed for complex, multi-step tasks. Notably, the model capability gap between these tools continues to narrow — tooling ecosystem, interaction experience, and pricing strategy are increasingly the deciding factors for users.

Looking Ahead

On the AI developments he's most excited about in the next 6 to 12 months, he pointed to two directions:

Mobile-first coding: Using smart glasses and better mobile visual feedback to "write code while walking, dispatch agents on the go" — breaking free from the occupational constraint of sitting at a desk for hours.
Native video multimodal input: Currently, video input relies on FFmpeg frame-extraction hacks that are slow and fragile. Native video understanding would dramatically improve scenarios like debugging animations.

He also predicted the trend of agents as independent entities — with their own credit card information, Gmail accounts, and even credit scores, acting on behalf of users across the internet. "Agent traffic will soon surpass human traffic. I think Stripe will be the first to create this kind of entity — which is also why they acquired OpenRouter."

On the security risks of open-source models, he's cautious: "It's already happening. Hugging Face had to switch to open-source models specifically because closed-source models refused security-related prompts." He believes big tech's concerns about model safety aren't all hype: "Sometimes they're just telling the truth."

The author cites Stripe's acquisition of OpenRouter as evidence for the "agents as independent entities" trend. OpenRouter is a platform that aggregates multiple AI model APIs, letting developers access different vendors through a unified interface. As the core infrastructure player in internet payments, if Stripe integrates model routing capabilities, it means agents executing payments, subscriptions, and commercial transactions can do so under a unified identity and credit system — without routing through a human account each time. This extends to the concept of an "agent economy": agents are no longer just tools, but independent digital entities with their own identities, credentials, and authorization to act — signing contracts, placing purchase orders, managing subscriptions. When agent-driven traffic surpasses direct human traffic, the entire existing internet infrastructure for identity verification, risk control, and bot detection will need to be redesigned. The implications for the broader web ecosystem go far beyond improvements in coding efficiency.