Level 1 of 3
Developers use code completion and occasional chat. The target is level 3, orchestrated agentic development. It can be reached without a rewrite. The path starts with building the safety net.
What is strong- The application framework is current. No large rewrite is needed.
- Money is handled correctly. On synthetic data, the core of the calculation gave the same results as an independent reference calculation.
- Most of the code is small: the median function is 13 lines long.
- High
There is no safety net. No working tests, no pipeline and no locked list of dependencies, so an agent cannot tell whether it broke something. On synthetic data we found three calculation errors that stay invisible without tests.
- High
Three serious security risks: one in sign-on, one in permission checks, one in an outdated library that processes what users upload. Each fix is small.
- Medium to high
Customer environments are not code and not under version control, and they are not isolated from each other. This has to be fixed before any AI feature reaches customers.
The gaps stay open and the calculation errors go unnoticed. Agents would copy the patterns already in the code, such as large functions and hand-written permission checks, far faster than a person would.
- Finding
- A permission check that never blocks. The check is a method, but it is used without being called, so the condition is always true. Accounts that the user interface restricts can read a register through the interface for integrations.
- Severity
- Medium. It needs an account, so an outsider cannot use it.
- Evidence
- Four call sites and the method's definition, each by file and line.
- Fix
- Four missing pairs of brackets, and a regression test that proves the fix.
| Phase | What it holds | The usual wayperson-days | With the methodworking days |
|---|---|---|---|
| 1 | Immediate fixes: eight small corrections before any other development | 10.5 to 18.5 | 3.1 to 5.5 |
| 2 | The safety net: locked dependencies, tests, a pipeline, instructions for agents | 21 to 40 | 4.6 to 8.7 |
| 3 | A new interface where permissions are decided in one place, with an audit log | 24 to 43 | 5.0 to 8.9 |
| Phases 1 to 3 | 56 to 102 | 13 to 23 |
Both columns are estimates. Person-days describe an experienced team working the usual way. The second column is building and testing with the method; it leaves out the client's own time. In the first two phases the safety net is still being built, so the gain is smallest there.
Why this orderAgents are only as good as the safety net of the code base. Without tests and a pipeline, an agent does not notice that it broke a calculation, and a person notices only when a customer calls.

