Choose a tech stack you can maintain long-term

Fast code generation is great but the architecture is where real commitment comes in. Compare architectures on representative changes, including testing, security, operations and migration costs.

Build & Validate
Evidence-led buildingAI & Automation

Adapted by the EWOR Team from Paul Müller’s article (opens in a new tab)2 min read

Choose the software you will have to operate with intention

Generating code can be quick. Living with the architecture is a much longer commitment. Paul Müller, an EWOR Partner and co-founder of Adjust, argues that a constrained, uniform tech stack gives AI agents fewer opportunities to go wrong. But that doesn't mean you should automatically adopt his preferred language or framework. Test the approach against the software your team actually needs to build and maintain.

Frame the decision around the software you must operate. What changes will agents repeatedly make? Which interfaces need to remain stable? Where would an unnoticed error be costly? A compiler, type system or test runner may give useful feedback, but its value depends on whether that feedback catches the failures your product actually faces.

Compare the same work in each stack

Pick a representative task and ask for comparable implementations in the stacks you are seriously considering. Include a routine change and a difficult one, such as modifying an existing interface without breaking its callers. Set the same acceptance criteria for each. Record whether errors were caught, how understandable the feedback was, how much human correction was needed and whether a subsequent maintainer could safely change the result.

Then inspect what the agent did not have to confront during the exercise. How will dependencies be reviewed and updated? Can the team debug production incidents? What does deployment require? Are there security, compliance or data-handling constraints? A stack that produces a quick passing demonstration may still impose unacceptable operational or migration costs. Conversely, a team’s familiarity with a stack has real value even if another looks simpler to an agent.

Write down the trade-off and the conventions

Document the trade-off rather than declaring a winner from code-generation speed alone. Specify preferred patterns, permitted dependencies, interface boundaries and required checks for the chosen stack. Revisit those constraints when the product or agent tooling changes. Consistent patterns make review and maintenance easier. They do not make generated code automatically correct.

Start with a small, reviewed comparison on a sanitised backlog task or in an approved environment. Check vendor data handling, repository permissions and secret exposure before supplying code. Keep the acceptance criteria the same, and let the maintainers assess security and maintainability before considering migration.

Source: EWOR article by Paul Müller (opens in a new tab)

Get the next signal before everyone else.

Frameworks and contrarian theses for outlier founders, straight to your inbox. No fluff, no recycled advice, weekly.

The EWOR Signals newsletter