Automating Myself Out of Development
A developer recounts their journey of incrementally automating their development workflow using Claude Code, evolving from interactive sessions to a daemonized, checkpoint-based process. This deep dive into AI agent orchestration, utilizing GitHub issues as a planning board, resonates on HN by exploring the practicalities and challenges of delegating core development tasks to AI. The story highlights the continuous shift in bottlenecks, from coding to meticulous review and strategic architectural oversight.
The Lowdown
Nune Isabekyan shares an evolving personal workflow of delegating development tasks to AI, specifically Claude Code, charting a course from initial interactive coding to a highly automated, daemon-driven system. The author emphasizes a pragmatic approach, acknowledging AI's limitations and the critical need for human oversight and trust at various stages.
- Initial Automation Challenges: Started with interactive Claude Code sessions for brainstorming and implementation, but quickly encountered context-switch fatigue and the burden of constant human validation.
- Securing and Isolating: Moved the AI agent to an isolated EC2 instance to reduce the 'blast radius' of potential errors, which unexpectedly revealed context leakage issues between projects.
- Checkpoint-Style Workflow: Realized the need for a non-interactive, scheduled process with clear human-AI handoffs, requiring persistent state, automated task pickup, and distinct 'stop' points for human review.
- GitHub as an Orchestrator: Leveraged GitHub issues with labels and comments to manage the AI workflow, allowing for phased execution and isolation of context for different AI sub-agents (e.g., brainstorm, spec, plan, implement).
- Daemonization: Implemented a
tick.shcron job to automate the process, picking up 'ready' issues, invoking Claude Code, and updating issue states, effectively shifting coding to background tasks. - Enrichment and Auto-Brainstorming: Introduced an 'enrichment' step where AI pre-researches issues to provide context, followed by an 'auto-brainstorm' pass that generates specs with confidence levels, maintaining human gates for review and approval.
- Remaining Bottlenecks: Identifies QA, architectural consistency, and intelligent feature categorization as the next significant challenges, acknowledging that while throughput increases, average code quality may not automatically improve.
The author concludes by stressing that AI should not be delegated thinking, and while the process increases throughput, it also heightens the importance of human code analysis, review, and architectural guidance. The journey is presented as an ongoing exploration of where the line of safe AI delegation lies.
The Gossip
AI Delegation & Trust in Code Generation
Commenters discuss the appropriate level of AI involvement in development, especially distinguishing between isolated components and complex architectural tasks. While some, like the author, maintain a highly supervised approach, treating AI as a 'mob programming participant' rather than an autonomous developer, others advocate for greater automation, claiming 'auto-mode is fine' and that current tools allow for spawning subagents for parallel tasks. The debate touches on whether AI is better suited for maintaining existing systems or building from scratch, with differing opinions on its efficacy for initial creative phases.
The Evolving Bottleneck: From Coding to Review
A significant theme echoes the author's observation that AI shifts the primary development bottleneck from writing code to thoroughly reviewing AI-generated output. Many commenters express the challenge of reviewing large pull requests or complex changes created by AI, often finding it more draining and time-consuming than writing the code themselves. This highlights a critical trust gap: developers feel compelled to meticulously review AI's work, especially for architectural implications, due to concerns about quality, correctness, and potential technical debt, making the 'review' phase the new hurdle.
Real-World AI-Assisted Development Experiences
Several users share their practical experiences in employing LLMs for coding, ranging from building entire products from the ground up to maintaining existing systems. Examples are provided of projects where a high percentage of the code was written by AI agents, sometimes with detailed logs of prompts and commits. This prompts discussion about the actual productivity gains from using AI, particularly for experienced developers, and the level of dedication required to achieve substantial results, with some suggesting that while feasible, it may not always be more efficient than traditional methods.