How AI Coding Agents Became Part of My Development Workflow
My experience with AI coding agents, from smaller tasks to planning, goals, and reusable skills, while keeping review and validation part of the development workflow.
When I started exploring AI coding agents more seriously, my first concern was figuring out how to make better use of these tools without spending unnecessary effort and tokens. Codex was my main starting point.
Over time, I saw the value of involving an agent at different stages of the work: investigating problems, understanding a codebase, planning changes, and reviewing the output. The quality of the result also depended on how I guided and followed up on those tasks.
---
From Smaller Tasks to Feature Development
AI coding agents can work with project files, run commands, and help with software engineering tasks, depending on the capabilities and permissions available.
Early on, I found value in smaller tasks: investigating bugs, consulting documentation, reviewing changes, and examining tests. That support also helped me improve process documentation and arrive at more reliable solutions.
As I became more familiar with the tool, I started asking for help with larger tasks, such as discussing architecture and developing features. Reviewing the generated code and validating application behavior remained part of the work. For certain tasks I carried out with Codex, this combination reduced my development time by around 60% to 80%.
That figure reflects my experience with those particular tasks. The outcome depends on the problem, the context provided, and the review needed before the work can be considered complete.
---
Matching the Agent's Settings to the Task
Depending on the tool, you can choose the model and reasoning effort. A higher level may help with complex tasks, but a quick question or a small change may not need the same depth of analysis.
It is also worth checking whether the agent has the information it needs. If a business rule, a relevant file, or a project constraint is missing, increasing the reasoning effort can still leave the problem poorly defined.
Permissions are another part of this setup. An agent may ask for approval before taking certain actions or operate more independently, depending on the environment. Understanding those boundaries helps you decide how to oversee the work.
Files, images, and other materials can provide additional context during a task when the tool supports them. Choose what helps explain the problem: the code related to an error, the expected behavior, or the documentation for an integration, for example.
---
Planning While Decisions Are Still Open
A feature can seem simple until we start thinking through the details. How should it behave offline? What happens if important data is missing? What if the user takes a different path through the workflow?
These questions help reveal when the requirements still need refinement. Agents such as Codex and Claude can support that discussion by raising questions and organizing a proposal before implementation.
In Codex, `/plan` is an example of a feature designed for this stage. The resulting plan can be reviewed and adjusted before execution begins. This process helps make assumptions explicit before they lead to rework later.
Once the expected outcome and constraints are defined, you can give the agent a specific goal and completion criteria. In Codex, `/goal` supports this kind of work. Other agents may offer different ways to organize execution.
The clarity of the requirements is central to this choice. Even a large task can be driven by a goal, provided there is enough information to carry it out and check the result.
---
Reusing Recurring Instructions
As you use an agent more often, some instructions start to repeat. In a code review, for example, you need to consider project conventions, error handling, and behaviors that must be preserved.
In tools that support them, these instructions can be organized into reusable skills. In Codex, they are described in a `SKILL.md` file and can be invoked with `$`.
Recording these instructions makes it easier to repeat a process without rebuilding all the guidance in every conversation. It also lets you refine what should be checked as gaps emerge in reviews.
Each task still needs its own context, though. A skill can guide the review, but you need to explain what is changing and which behaviors are expected.
---
Following the Work Through to Completion
Over time, I started involving the agent in larger parts of development. Combining planning, implementation, and review helped me deliver more robust solutions in less time for certain tasks.
To assess that improvement, I consider the work required to bring a task to completion. That includes checking the modified code, validating application behavior, and making any necessary corrections. A solution that looks appropriate in isolation may still fail to meet a project constraint.
The time saved also made room to study software engineering, processes, and artificial intelligence, as well as identify gaps and explore better approaches to the solutions I develop.
Today, I try to identify where an agent can contribute and which criteria I will use to check the result. That approach guides both my use of Codex and my exploration of other AI coding agents.