In software development, the division of labour is shifting. It is no longer just the writing of code that is delegated to AI tools, but also what comes afterwards: formulating tests, reviewing changes, naming vulnerabilities. In a podcast conversation, Richard Seidl and Benedikt Stemmildt talk about agents that improve themselves – and about why Codex is sometimes the best reviewer for Claude Code.
This observation is unspectacularly phrased and yet it is the interesting point. Because it describes a constellation that is new in everyday project work: two different systems do not work side by side, but on top of each other. One produces a proposed change, the other reviews it.
Why a second model finds anything at all
A code review thrives on someone looking in with a different perspective. Anyone who has designed a solution themselves carries that solution's assumptions with them – and therefore systematically overlooks certain errors. This applies to humans, and apparently in a similar form to AI models: an agent that reviews its own output remains in the same line of thinking.
This is precisely where the practical value of different tools lies. A different model, trained differently and tuned differently to tasks, brings different priorities. It may comment on what the first model took for granted. That is no miracle cure, but it is more than a double-click on the same button.
Self-improvement is not automatic
The second term from the conversation – agents that improve themselves – sounds bigger than it initially is in everyday project work. At its core, it means a feedback loop: an agent produces something, receives feedback, adjusts its next attempt. This feedback can come from tests, from a build, from a linter or indeed from a second agent.
What matters is that this feedback loop has a reliable signal at all. Without automated tests, without a clear definition of "done" and without a reproducible environment, the loop spins in a vacuum. An agent that reacts to unclear feedback produces more changes – not better ones.
For projects, this means: the prerequisites for meaningful agent work are pretty much exactly the same ones that make for good software development anyway. Proper test coverage, a working continuous integration pipeline – that is, the automatic build and test run on every change – and clean boundaries between modules now pay off twice over.
What clients notice about this
From the client's perspective, the technical constellation is less interesting than its effect. That lies above all in the lead time. A review that previously waited for a second person to become available can now happen in part immediately. Small issues – forgotten error handling, sloppy naming, missing test cases – are found before a human opens the ticket.
This shifts the role of human review upwards: away from formalities, towards architecture, functional correctness and questions that can only be answered with knowledge of the business model. That is the real efficiency, not the saving of people.
At the same time, a new risk arises. If changes are produced faster and waved through faster, the amount of code that nobody has really read grows. An agent review is an additional net, not a substitute for responsibility. Whoever approves is liable – technically, contractually and towards their own users.
What makes a clean setup
Some guidelines can be derived from practice that apply regardless of the specific tools used:
- Separate roles. The agent that produces a change should not be the same one that gives it the final assessment.
- Automated checks first. Tests, build and static analysis run before every model review. What a machine can check deterministically does not need a language model to assess it.
- Anchor human approval. For changes to interfaces, data models, permissions and payment processes, the decision remains with a person.
- Ensure traceability. Which agent gave which hint, what was adopted, what was discarded? Without a log, the basis for later troubleshooting is missing.
- Clarify data flows. Which source code leaves your own system and goes to which provider? This question belongs before the tool selection, not after it.
Tools come and go
The names in this discussion will change. Which model is considered the better reviewer for which task today may look different in a few months – and that is precisely why it is worth not optimising for a single product.
What remains stable is the structure: defined steps, interchangeable tools, clearly assigned responsibility. A team that has set up its development process this way can replace an agent without reinventing the workflow.
For web and software projects, this means: the benefit of AI agents in review depends less on the model than on the quality of the pipeline into which it is embedded. Those who have set up tests, automation and approval paths properly gain speed without giving up depth of review. Those who have not merely accelerate the production of unreviewed code – and only notice that in operation.
Sources
- Software Testing: Mit Claude und Codex Software auf Steroiden entwickeln — heise developer News