Presenters
Source
The AI Co-Pilot Revolution: Coding, Collaboration, and the Future of Engineering 🚀
We’re living through a seismic shift in the tech industry, and the way we write and think about code is changing at an unprecedented pace. In a recent fireside chat, Jesse Chen, Director of Product on the Dev Infra team at Meta, sat down with Boris Cherny, Head of Cloud Code, to dive deep into this transformation. From AI writing 100% of code to the rise of “loops” and the future of collaboration, this conversation offered a fascinating glimpse into what’s next.
The AI Code Generation Tipping Point 💡
The room buzzed with excitement as Jesse Chen kicked off the discussion by polling the audience on AI code generation. A significant portion raised their hands when asked if they use AI to write code. But the real eye-opener? When asked if 100% of their code is AI-generated, many more hands shot up than before, signaling a profound industry change.
Boris Cherny himself is a testament to this shift. He revealed that 100% of his code has been written by Cloud Code since November of last year, a statistic that continues to astound. He even shared that since March, he’s used a staggering 8 billion tokens, highlighting the immense scale of AI-assisted development. In a surprising twist, he confessed that most of his coding now happens on his phone, a notion he would have dismissed just months prior.
Efficiency vs. Innovation: Balancing the AI Budget 💰📈
The conversation then pivoted to a critical concern for many organizations: the balance between the increasing cost of more capable AI models and the need for demonstrable efficiency and ROI. Companies are exploring strategies like setting monthly budgets of $1,500 per engineer.
Boris emphasized that framing this through an ROI lens is crucial, rather than just focusing on cost. He advocates for broad experimentation, encouraging companies to provide tokens to everyone, not just engineers, but also product managers, designers, and data scientists. This democratization of AI access, he argues, can unlock innovation from unexpected corners of an organization.
Once successful internal use cases emerge, the focus shifts to cost control and optimization. Cloud Code offers solutions like per-seat cost controls, advisor models, and departmental budgets.
Measuring True Return: Beyond Lines of Code 📊
The traditional metrics for AI’s impact, like percentage of code written by AI or increased lines of code, are becoming outdated. With AI now writing 100% of code for many, the question becomes: how do we measure the return?
Boris highlighted that productivity improvements are now in the hundreds of percentage points, with Anthropic seeing an 8x increase in code per engineer since the start of the year. The real measure of ROI now lies in understanding how code output per engineer is accelerating and identifying other bottlenecks in the development lifecycle, such as idea generation and go-to-market speed.
The Rise of “Loops”: Agents Prompting Agents 🤖➡️🤖
The discussion ventured into the exciting realm of “Loops,” a concept that might define the next hype cycle or signal a fundamental shift. For engineers, Loops can be understood as a higher-order function, an abstraction layer above agents writing code. The progression is clear: source code written by hand → agents writing code → agents prompting agents that then write code.
Boris shared a personal anecdote: he now uses Loops for tasks like code reviews and monitoring feedback on Threads. He estimates that around 30% of his code is written by Loops daily, with the potential to reach 100% on certain days. This evolution signifies a move towards more autonomous AI systems.
Co-Work: Empowering Non-Engineers with AI 🛠️
Anthropic’s investment in Co-Work was a key topic. Co-Work, essentially Cloud Code for non-engineers, leverages the same powerful Claude agent SDK but with enhanced guardrails. It operates within a virtual machine, with OS hooks to prevent accidental data deletion and robust prompt injection protection.
Boris shared compelling use cases for Co-Work beyond coding:
- Project Management: Automating stand-ups by having Co-Work poll engineers for status updates and populate spreadsheets.
- Travel Booking: Co-Work autonomously books flights and hotels based on calendar events and email confirmations, even handling multi-leg international trips without human intervention.
The magic of Co-Work lies in its ability to orchestrate and combine various tools, offering a revelation akin to experiencing a chat app for the first time.
Fable: The Next Leap in Model Capability ✨
The conversation touched upon Fable, Anthropic’s latest advanced model. While access has been limited, Boris described Fable as a leap of at least the same size as the jump to Opus 4.5, potentially even larger. He lauded Fable’s nuance, dimensionality, and ability to grapple with complex problems, comparing its thinking to that of his smartest co-workers.
Fable excels in areas like data analysis, requiring deep “why” questioning, and debugging, where it can form hypotheses and chase down evidence effectively. Boris confessed he “ran out of hard problems to give it,” with Fable consistently delivering with minimal prompting.
Model Stratification: Choosing the Right Tool 🎯
With a family of models available, the question arises: when to use which? Boris’s approach is pragmatic: he uses Fable for everything, even acknowledging the potential cost. He believes the opportunity to increase return with Fable likely outweighs the cost savings from using less advanced models. The focus should be on maximizing the return by leveraging the most capable models.
Tackling Bottlenecks: From Code to Collaboration and Beyond 🌐
The core theme resonating throughout the discussion is the relentless pursuit of solving the next bottleneck.
- Collaboration: While GitHub exists, the need for better collaboration tools integrated with AI code generation is clear. Cloud Code is working on solutions, and in the interim, integrating with Slack, Teams, or G Chat via MCP is recommended.
- Engineering Focus: With AI handling coding, engineers are shifting their focus to idea generation, market research, customer interaction, and strategic alignment. Prompting the AI effectively and identifying the “next thing” become paramount.
- Code Review & Security: To address the surge in AI-generated code, Anthropic developed Claude Code Review and Claude Security. These products, built internally and now available externally, aim to fully automate code review and proactively scan for vulnerabilities, catching issues even human pen testers might miss.
- CI/CD Optimization: Boris shared an example of prompting Claude Code to optimize CI timings, resulting in a 50% reduction in CI time through four pull requests, a task that would have previously taken days or weeks.
Workflows and Loops: The Future of AI Orchestration ⚙️
The introduction of dynamic workflows and loops represents a significant advancement. Workflows allow Claude to orchestrate dozens, hundreds, or thousands of sub-agents dynamically, a new form of test-time compute. This enables Claude to tackle complex problems by breaking them down and assigning them to specialized agents.
The Long Game: Maintenance and Long-Term Code Health 🌳
Maintenance, often a bigger challenge than initial coding, is also being addressed by AI. Boris is experimenting with Loops for maintenance tasks, such as:
- Improving code base architecture.
- Identifying and fixing flaky tests.
- Deleting redundant tests.
- Unifying duplicated abstractions.
He trusts the AI to handle these tasks, reviewing pull requests after the loop has executed, especially with the latest models and the “use a workflow” command, which significantly enhances output quality.
The Vision for Claude Code: Ubiquitous, Capable, Experiential 👁️🗨️
Anthropic’s vision for Claude Code over the next year centers on three key pillars:
- Most Capable Agent: Continuously improving AI’s ability to handle long-running tasks, enhance code security and quality, and better align with user intent.
- Works Anywhere: Seamless integration into existing team workflows, without requiring users to switch to a separate platform.
- Experiential Capabilities: Developing products that allow users to fully experience the power of new models, much like Claude Code did for the leap in coding capabilities with Opus 4.5.
Overcoming Laziness: Empowering Engineers Through AI 🧠
A critical question remains: how to prevent engineers from becoming complacent and blindly accepting AI output? The answer lies in two fronts:
- Enabling Safe Automation: Instead of relying solely on human oversight, the focus is on making Claude do the right thing safely. The introduction of Auto Mode in Claude Code, where the model makes permission decisions based on conversation context, drastically reduces “prompt fatigue” and enables longer-running agents. This is made possible by Claude’s advanced resistance to prompt injection, achieving a success rate of around 1%.
- Fostering Continuous Learning: Tools like exploratory and learning output styles in Claude Code are designed to keep engineers engaged and learning. These styles provide detailed explanations of architecture, languages, and code snippets, effectively turning AI into a powerful educational tool. This ensures engineers remain informed and adaptable, even as their tech stacks evolve.
The journey of AI in software development is far from over. As we continue to solve bottlenecks and push the boundaries of what’s possible, the role of the engineer is evolving from creator to orchestrator, curator, and strategic thinker. The future is not about replacing humans, but about augmenting their capabilities to achieve unprecedented levels of innovation and efficiency.