Beyond text
Enabling visual review of AI generated code bases
Everybody loves doing code reviews 😉
Doing code review is like going to the dentist. You rarely looking forward to it, but you know the consequences of not doing it regularly, so you do it out of reason. Till now, it probably did not took a lot of time. Research says 1-5 hours a week (link, link, link) Unfortunately (or not) AI coding agents flipped this ratio. Now we now longer write that much code, now we mainly review it. I don’t know about you, but for me looking at text diffs for several hours a day is not a dream come true.
There’s some deep truth in this post, because very often what we end up doing is skipping the review or just skimming through the code and giving it a “looks good” approval. The result is well described in the post by Rob Zuber, CTO at CircleCI:
and the Cognition Team
A picture is worth a thousand words
At Noesis we believe that visual code review is the answer. I know what you might think - there were multiple attempts to visualize code, and they did not delivered. They did not, because the only thing they did, was to convert code to its visual representation, one to one without eliminating any complexity. Tools that are actually useful for developers to visualize their system, are the ones that introduce abstraction reducing cognitive load (UML diagrams, C4 model tools). Unfortunately none of those tools produces visualizations automatically. This is where Noesis comes in. We were able to combine the best of both worlds by automatic detection of building blocks, that people use to reason about their code base.
How does Noesis work?
First you need to tell Noesis what building blocks you are using to reason about your code (entities, services, pipes, filters, aggregates e.t.c.). For each building block, you define convention, that will be used to locate it (name pattern, annotatedWith e.t.c), and that is the end of the setup.
After that you can start coding with your favorite agent. After your agent is finished generating code, you ask Noesis to scan you your project, and you can verify, in the blink of an eye, if all expected building blocks and relation are there.
Noesis uses static code analysis together with already defined conventions, to produce P3 model - simple, but powerful software model we designed (more in the future articles). Based on it we can create a multiple types of diagrams in fully deterministic way.
After double clicking on diagram elements, you can verify if all business rules and data elements are there and how the interactions are looking on sequence diagram.
Text descriptions above are LLM generated, based only on code that is related to particular building block. P3 model works as knowledge graph based RAG for code base and enables precise context management and mitigates hallucinations.
If you want to verify domain module structure, we have a view for that too
With a couple of clicks you can quickly review what was done, and verify if it makes sense. If you spot any error you can dive into code knowing what you are looking for. Better code understanding allows you to quickly correct agent course by remarks like: “Add missing event”, “Create a common service for a given functionality to avoid duplications”, “Move the logic from a controller to a service” or “from service A to service B”
You can keep adding steering rules for agent to avoid future flaws, but our practice shows that they are not always used. Having the regular reliable, visual validation of code changes gives the so much needed additional level of control.
On our roadmap
This is only a beginning, we have a lot of ideas, that will make working with coding agents on large codebases a breeze.
Diffs
Enable comparing revisions, so that you can better see what has changed since last commit.
Scenarios
Display informations about scenarios (automated tests) for each element, so that you can see which rules are tested, and what coverage you have
Architecture view
Add more technical view, with layers, ports, adapters, API, e.t.c.
Service Map view
Add service map view for distributed systems, where you will be able to see all dependencies and messages that flows between your services
Spec visualisation
In Spec-Driven development thorough review of design.md can also be time consuming. We are working on Claude Skill that will convert specs into P3 model, and therefore enable visualising it before implementation starts.
But AI agents can already do it
Yes. AI agents are becoming better and better at exploring and explaining codebases, but should they be judges in their own case? It’s safer to use deterministic tools to verify generated code. We already do it with agents running compilers, automated tests, code quality tools, architecture tests e.t.c. Noesis fits in this category of tools. Tools than can be trusted; tools that produce results without hallucinations. That’s why we rely on hybrid solution (static code analysis for locating building blocks and LLM for describing their insides) instead of pure AI-based approach.
SEE it for yourself
If you like to play around with our demo instance, it is available here.
If you are interested in trying it out on your project, join our beta here
Visit our website for more (Discord, newsletter, documentation etc)