Building Better Developer Workflows with AI
I had the opportunity to sit down with Chris Traganos, Head of Developer Evangelism for Alexa at Amazon, to talk about shifting AI from coding assistant to comprehensive development tool, covering the 84% of developer work that isn't writing code.
Key Takeaways
- Target the 84%: Developers spend only 16% of their time coding. AI applied solely to that slice explains why many teams see minimal ROI.
- Context over prompts: Getting good results requires treating AI like a teammate. Provide deep context and collaborate on a plan before asking for execution.
- Local agents unlock automation: On-device agents with file system access can configure development environments, manage tmux sessions, and handle setup tasks that would otherwise consume hours.
- Side projects build skills: In an AI-driven market, curiosity-driven programming keeps skills sharp and maintains employability better than certification chasing.
- Universal React Native: Expo and React Native Web enable true cross-platform development from a single codebase across iOS, Android, web, and TV.
The 84% Problem
Most AI productivity discussions focus on code generation. GitHub Copilot, Cursor, autocomplete suggestions. These tools optimize writing code faster. The problem is that coding represents roughly 16% of a developer’s weekly work.
The remaining 84% disappears into technical overhead. Writing documentation. Communicating with teams. Reviewing pull requests. Managing tickets. Creating technical specs. Updating Jira. When companies report disappointing AI returns, the explanation often traces back to applying AI tools only to the small coding slice while ignoring everything else.
Practical applications for the 84% include using AI agents to draft technical requirement documents from vague ideas, automating Jira ticket creation by teaching the AI to interface with the Atlassian API, and breaking down technical plans into actionable, formatted tasks. These workflows target the overhead that actually consumes developer time.
Developing an AI Palate
A significant barrier to AI adoption is what could be called “AI friction,” the awkward phrasing, cliché outputs, and hallucinations that characterize early LLM interactions. Many developers try AI tools, experience these issues, and conclude the technology isn’t ready.
Effective AI use resembles learning to cook more than memorizing prompt templates. The skill isn’t about “prompt engineering” as a rigid science. It’s about developing a talent for providing context and guardrails, similar to communicating with a human teammate who has vast knowledge but no project context.
“Vibe writing” plans alongside an agent represents one effective approach. The developer acts as orchestrator or guide rather than simple prompter. Instead of asking AI to generate code directly, the collaboration produces a shared plan that both parties refine before execution begins.
On-Device Agents and Local Access
A distinction exists between using AI through a browser interface versus running agents directly on your machine. Browser-based AI operates within the sandbox of your conversation. On-device agents can touch your local file system, opening entirely different automation possibilities.
Tmux and terminal configuration offer a practical example. Setting up tmux properly, configuring vimrc files, learning keyboard shortcuts: these tasks are tedious enough that many developers never complete them. An on-device agent can handle this setup work, finally enabling developers to learn and configure tools they’d been avoiding.
The parallelization benefit proves equally significant. With tmux, a developer can manage five or six side projects simultaneously. Set an agent to work on a bug in one terminal pane, then switch to another session to continue work elsewhere while the agent processes. This workflow treats AI agents as background workers rather than interactive assistants requiring constant attention.
Career Longevity Through Play
The current tech economy has compressed certain roles. Tasks that junior developers once handled are increasingly automated. Basic CRUD operations, simple bug fixes, and boilerplate generation all fall within AI capabilities.
The response isn’t to chase certifications or desperately upskill. Instead, treating programming as a form of play keeps skills naturally sharp. Side projects become the equivalent of video games, an escape and source of intellectual engagement that happens to maintain employability.
Following curiosity rather than career planning tends to lead developers toward interesting problems. Those interesting problems build exactly the skills that remain valuable as automation handles routine work. The developers who program because they find it genuinely engaging grow automatically while others scramble to stay relevant.