ChatGPT
CodexHow I use AI-assisted investigation, implementation, and verification in my practice, and what agentic coding is changing across software development.
Agentic AI coding uses an AI system that can work through a sequence of development actions: inspect a repository, reason about a task, propose or make edits, run permitted commands, and use the results to revise its next step. Its usefulness comes from that working loop, beyond producing a standalone code suggestion.
A developer still needs to define the goal, provide relevant context, set the scope of access, and decide whether the result is appropriate. The agent’s tools and permissions determine what it can inspect or change. Tests and build output provide evidence about the work, but neither removes the need to assess the requirements and review the changes.
I have embraced AI as part of my practice because I want to keep improving how I build software. It helps me explore implementation options, work through repetitive changes, and bring ideas into a form I can inspect and refine. I pair those capabilities with my engineering experience and a clear definition of the result the business needs.
An agent can help apply a bounded pattern across related files, connect a page to its data, or prepare an implementation for local review. Clear instructions and an existing codebase give that work useful constraints.
Business impactPotential time savings can create more room for product decisions and refinement. I assess the complete task, including prompting, waiting, review, and rework, instead of treating generated code as delivered value.
Agents can assist with tracing related code, identifying callers, drafting checks, and running the commands selected for the change. A concrete input and expected result make the work easier to inspect.
Business impactUsed carefully, this can support more consistent verification and help surface omissions. An agent can also misunderstand the system or write checks that merely agree with its own implementation, so independent review matters.
AI assistance makes it practical to draft alternatives and review them in a working environment. The developer can compare the result against the product’s audience, business requirements, and existing architecture.
Business impactMy aim is better products through more informed iteration. The benefit comes from choosing and refining the right implementation, with understandable code and support requirements.
Agentic tools are changing where effort is spent: defining tasks, supplying context, reviewing changes, and validating behavior become more visible parts of delivery. Industry reports show adoption, but the evidence does not support one universal productivity multiplier. The findings below describe different tools, dates, populations, and measures.
GitHub’s 2026 Octoverse report documents rising agent participation in pull requests and substantial growth in platform activity. This is evidence about workflows on GitHub; more commits or pull requests alone do not establish better software or prove that AI caused the growth.
GitHub Octoverse 2026Anthropic’s June 2026 analysis of Claude Code sessions describes people making much of the planning decisions while the agent handles much of the execution. It is vendor research about one tool’s users, not a controlled measurement of every developer or a forecast that engineering judgment is becoming unnecessary.
Anthropic: agentic coding and expertiseMETR’s early-2025 randomized study found that 16 experienced open-source developers took 19% longer with the tools studied on 246 tasks. Its February 2026 follow-up reported indications of improvement with newer tools, but warned that selection effects made the current estimate unreliable. The lesson is to test the workflow in its actual setting rather than generalize either result to all coding work.
METR: 2026 productivity-study update and earlier resultsSoftware work includes understanding an existing system, making coordinated changes, and checking that the result fits the business. Writing the first version of a function or page is one part of that responsibility.
I have embraced AI to help with that process. It can take on investigation and repetitive implementation tasks while I guide the requirements and evaluate the result. Thinking ahead means learning how to use new tools well and improving the delivery process as those tools develop.
I describe the intended behavior, the audience, and the constraints before asking an agent to make a change. During this website work, clarifying what belonged under Portfolio and Case Studies determined the implementation.
The agent can inspect existing layouts, data, styles, and navigation behavior. Working with those conventions helps keep a change connected to the rest of the site.
A local preview, an inspectable change, and relevant checks give me something concrete to evaluate. I can then adjust the content, interaction, or scope against the original intent.
I see adopting AI as an ongoing part of improving my practice. I want the tools to help me build clearer, more useful products, while maintaining responsibility for what is delivered and how it will be supported.
My aim is to make room for better decisions and more useful iteration by reducing repetitive effort. The value comes from combining the tools with experience, a clear product goal, and verification that addresses the real behavior being changed.
Tell me what you’re building or what needs fixing.
Start a conversation
