Most teams work in systems that have grown over years, with data, interfaces and rules that nobody holds in their head completely anymore. Coding agents speed up the work there too. In an existing system, though, they produce errors that surface late: a second implementation next to the existing one, or a ticket that describes only half of the business situation.
This series describes my attempts to organise that work. The original agentic engineering pipeline proposed fixed roles and a quality gate at every handover. Later posts examine the flow using experience from two internal projects. They also show where additional rules and handovers made implementation harder.
The revised flow starts with a small executable use case. Useful experimental code is extended after review. Independent acceptance checks the agreed criteria against the product, with further checks based on risk. Repairs count towards total effort. Previously accepted features should still work after several extensions.
Tickets, commits and run records provide the evidence for these reports. Research findings sit alongside my observations, with their respective limits. The posts below appear chronologically. The opening article explains the first design. The report on stagnation despite extensive testing explains the latest changes and the planned comparison of the procedures.