This write-up will outline the development workflow and explore why quality does not need to be compromised for speed.
Planning
Clear and precise front-loaded plans are the key to success, whether for a full product build or a feature added to an existing system. The reason is simple: staying adaptable during the process can only go so far. Adaptability is generally a strong trait to have, but the need to adapt during development can often be reduced through strong, well-thought-out planning.
However, the approach is vastly different depending on whether a product or foundation already exists. Building on top of an existing system can be a delicate process. Quality becomes heavily tied to what preceded it, and the correct architectural adaptations become paramount. Without adequate time spent assessing and considering the existing system during planning, poor delivery becomes far more likely.
Building from scratch creates an open canvas while leaving possible directions and approaches entirely ambiguous. Rushing this initial process, no matter the task, is a massive mistake that, more often than not, comes at a cost later.
Tools and context
The world of software engineering is evolving at a pace that is unusual even when compared with almost any other domain. One year ago, even one month ago, the space looked massively different. This is where adaptability becomes increasingly valuable, as a new 'revolutionary' tool or technology emerges every week and separating the signal from the noise becomes increasingly difficult.
Coding agents have become a central part of this approach to building software. However, like anything, there is a skill curve with these tools. Claude Code cannot simply be given to someone who has no idea how to design a system with the expectation that a functional and robust system will be created.
A fundamental, and potentially the most important, concept in AI-directed engineering is managing context coherently. This includes ensuring AI implementers have the correct plans, handovers and context for the system they are entering. The internal context system, Cortex, which is referenced separately, streamlines this process. A cumulative as-built record has also become very useful for handovers.
Letting agents run autonomously can be daunting. An as-built record captures any decisions made by the agents and tracks a coherent record of the build process itself. This allows later agents to understand why something was done, what the context looked like at the time and whether anything has changed since.
Model Context Protocol integrations or custom-built connectors further streamline the process, allowing agents to integrate directly with systems. Computer use provides a useful way to handle tasks outside what direct connectors can offer.
Implementation
Implementation is obviously crucial, but it becomes a direct result of what comes before and after. Poor planning, tools, context management, orchestration or auditing will produce poor implementation. More often than not, if all the moving pieces surrounding the implementation are configured correctly and given adequate time and diligence, then the implementation will follow.
Orchestrating the implementation, staying on top of delivery and spotting mistakes remain instrumental to the process. Confidence can be earned by being diligent and disciplined in the surrounding work.
Predefined workflows, task separation and allocation all become relevant. One of the strongest ways to keep implementation quality high is to avoid having the same agent handle multiple unrelated tasks. One of the simplest and most effective ways to deliver code is through sub-agents.
For example, one main agent supervises the work, almost like an on-duty manager. Spawning sub-agents to implement predefined, planned tasks gives each agent fresh context scoped specifically to what needs to be achieved. The more clearly defined the task, guardrails, quality standards and repository practices, the better the results will be.
The workflow articulated above is the approach currently in use. However, as mentioned earlier, the industry is moving at such a crazy pace that, ultimately, the approach used today will change.
Quality checks
With advances in what AI can achieve and deliver, verifying functionality is no longer an entirely manual task. This brings verification directly into implementation and ties the two together, with a final, full end-to-end check also forming part of the process.
This does NOT remove the need for manual verification or test fixtures. They become increasingly relevant, setting a strong foundation for any work that comes later. One increasingly useful approach is to use a completely fresh, separate auditing agent. This usually follows a single task or sequence of tasks and is tightly scoped to review the source code against predefined criteria. These usually include, but are not limited to, logical correctness, code quality and simplicity, security holes, functionality issues or bugs, and over-engineered solutions.
This catches many problems in their tracks and saves a lot of refactoring or similar work. Review is not limited to implemented code: collaboration with agents also includes reviewing design and architectural decisions, creating an ongoing discussion and review of the creative process.
Speed and quality
A final point of interest, as AI tools have progressed in quality and the level of trust that can be placed in them, is the almost ingrained notion that high quality and fast delivery cannot exist in parallel. The established belief seems to be that speed must sacrifice quality, or vice versa, though that perspective is shifting in real time.
A large number of people label AI-generated work as "slop", and a large proportion of it is. Another large cohort states that AI systems will take all jobs and leave everyone without purpose. These are two opposing viewpoints, and neither provides a complete picture in practice. AI systems can allow people who were already capable of delivering high-quality work to do it at a pace that previously seemed unreasonable.
The view here is that AI tools do not inherently replace humans, but can improve both the speed and quality of delivery across many fields in the hands of people who are adept enough to use them or learn exactly how they operate.