AI coding models and agents are now extremely capable. In many workflows, they are about 80–85% of the way there.
That sounds great, but the remaining 15–20% matters a lot. If you do not supervise closely, small mistakes compound.
What Goes Wrong
Agents still make too many assumptions. They often do too much in one pass, consume too much context, and then lose track of earlier decisions.
Once context gets compacted, they can recreate work that already exists. As context grows, they may ignore project rules, skip skills, and miss MCP-based context they should be using.
My Current Approach (April 2026)
Before I start an agent, I think through the task first. Then I break it into smaller, bounded tasks with clear goals.
During execution, I review generated files continuously. I push the agent toward reusable components, classes, traits, and functions instead of one-off code.
I usually run one execution agent at a time so I can supervise in real time. I may use separate agents for planning, but I keep build execution tightly controlled.
Workflow Rules I Follow
- Break work into small tasks.
- Review every generated file.
- Prefer reusable architecture over quick patches.
- Use git heavily.
- Revert bad changes quickly.
- Run experiments in separate branches.
AI-assisted development is very effective, but only with active engineering judgment. The tools are powerful; the developer still has to steer.
