The Manual Coding Retreat: Why One AI Engineer is Coding Without LLMs for Three Months
Miguel Conner, an experienced AI engineer from Aily Labs, has embarked on a three-month coding retreat in Brooklyn, New York, to focus on programming without the heavy reliance on AI tools. Despite his background in building AI agents and knowledge graphs, Conner argues that manual coding serves two critical functions: expressing intent and deeply learning a codebase. Having spent six weeks on this retreat as of March 2026, he reflects on the transition from using state-of-the-art models like DeepSeek R1 and Llama 3 to the traditional 'hand-coded' approach. This experiment comes at a time when many in the industry suggest that programming is a 'solved problem' due to the rise of AI agents and automated workflows.
Key Takeaways
- Intentional Disconnection: Miguel Conner is spending three months coding 'the old way' to rediscover the nuances of the craft.
- Deep Industry Experience: The author previously led projects at Aily Labs, building web search agents and knowledge graphs long before major industry releases from Anthropic and OpenAI.
- The Dual Nature of Coding: Manual coding is identified as a process of both writing desired logic and actively learning the underlying codebase.
- Contrarian Timing: This retreat occurs in early 2026, a period where many successful programmers claim that AI has effectively solved the problem of programming.
In-Depth Analysis
From AI Pioneer to Manual Practitioner
Miguel Conner’s decision to step away from AI-assisted development is particularly notable given his professional pedigree. At Aily Labs in Barcelona, he was at the forefront of the AI revolution, developing internal web search agents in early 2024—months before Anthropic published its influential 'Building Effective AI Agents' article and a full year before OpenAI’s DeepResearch. His work involved leading journal clubs to dissect the architectures of open-source models like DeepSeek R1, Ai2’s Olmo 3, and Meta’s Llama 3. This deep technical understanding of how LLMs are built and trained provides a unique perspective on why one might choose to temporarily abandon them.
The Hidden Costs of AI Agents
While using coding agents like Cursor and various LLMs, Conner identified a significant shift in the development process. He posits that traditional coding 'by hand' involves two simultaneous actions: the expression of what the programmer wants to create and the cognitive process of learning the codebase. The analysis suggests that while AI agents can handle the 'writing' aspect efficiently, they may disrupt the 'learning' aspect. By spending three months in Brooklyn focusing on manual input, Conner aims to reclaim the element of the craft that requires a deep, unmediated connection with the code, challenging the contemporary narrative that programming is a solved problem.
Industry Impact
This narrative highlights a growing tension within the software engineering industry as of 2026. As AI agents become more sophisticated, there is an emerging debate regarding the loss of 'codebase intimacy' and the long-term effects on developer expertise. Conner’s retreat serves as a case study for the 'craftsmanship' movement in software, suggesting that even as SOTA (State-of-the-Art) models become more capable, the human element of understanding and learning through manual labor remains a vital component of high-level engineering. It raises questions about whether the efficiency gained by AI comes at the cost of deep architectural comprehension.
Frequently Asked Questions
Question: Why did Miguel Conner decide to start a coding retreat in Brooklyn?
Conner moved to Brooklyn for a mix of personal reasons and a professional desire to focus on coding without AI for three months. He wanted to explore the 'old way' of programming at a time when the industry increasingly views coding as a solved problem.
Question: What was Conner's experience with AI prior to this retreat?
He spent two years at Aily Labs building AI agents, including early web search tools and knowledge graphs. He also led a journal club focused on the training and tradeoffs of models like Llama 3 and DeepSeek R1.
Question: What does Conner believe is lost when using a coding agent?
He suggests that manual coding allows a developer to learn the codebase while writing. Using an agent often focuses only on the output, potentially bypassing the deep learning process that occurs when writing code by hand.
