A Discussion on Self-Evolving SaaS Systems

GPT-6 Astra was just released recently, and people are saying it’s the generation closest to AGI so far.

Does that mean AI won’t just be able to write code in the future, but potentially automate the entire software development lifecycle?

I recently added MCP to Suggix, and it got me thinking about this.

AI is already incredibly capable, but it still needs humans to tell it what to do. In other words, it doesn’t have any intrinsic motivation.

But what if a SaaS could expose its user feedback data to an Agent, and use that feedback as the motivation for what the Agent does?

All you would need is a scheduled task that periodically pulls the latest user feedback, and the whole process could potentially run automatically:

User submits feedback (bug or feature request)
AI calls MCP to analyze the feedback, based on votes or its own judgment of how urgent/important it is
AI updates the code based on the user’s needs
AI runs tests, deploys the update, and notifies the users
The system keeps iterating based on subsequent user feedback

If this actually works, the entire SaaS iteration cycle could eventually run without the developer being involved at every step.

And more importantly, the product would be evolving based on what users actually need, rather than what the developer thinks they need.

I’m not sure how far this can really go yet, but I think the idea of a self-evolving SaaS is pretty interesting.

FAQ

Q1. Can AI really decide what a SaaS should build?

Potentially, but this is harder than simply reading votes. An Agent would need to consider user feedback, frequency, severity, customer impact, implementation cost, and the existing product context before deciding what should be built.

Q2. How can we prevent AI from making bad code changes?

The Agent shouldn't be allowed to directly modify production. A safer approach is to let it create a branch, implement the change, run tests, deploy to staging, and only promote the change after predefined checks pass.

Q3. Is exposing user feedback to an AI Agent safe?

Not automatically. User feedback should be treated as untrusted input because it could contain sensitive information or even prompt-injection attempts. MCP can provide access to the data, but the application still needs strict authentication, authorization, and input isolation.

Q4. Should AI have access to production?

Probably not unrestricted access. The Agent should receive only the permissions it needs, such as reading feedback, creating code, running tests, or deploying to staging. Production access should be limited by explicit policies and safety checks.

Q5. How does AI know whether its change actually solved the problem?

Automated tests are only part of the answer. A self-evolving SaaS would also need to monitor error rates, performance, usage, and subsequent user feedback to determine whether the change actually improved the product.

Q6. What happens if AI breaks production?

A self-evolving system needs automatic recovery mechanisms such as versioned deployments, feature flags, health checks, and rollback. The goal isn't to make AI incapable of making mistakes, but to make those mistakes contained and recoverable.

Q7. Could AI eventually automate the entire SaaS development lifecycle?

I think large parts of it are technically possible, but the hardest part may not be writing code. The real challenge is deciding what to build, validating that the decision was correct, and knowing when the Agent should stop and ask a human.

Q8. What is the biggest challenge for a self-evolving SaaS?

The biggest challenge is closing the loop safely: **feedback → decision → implementation → verification → deployment → monitoring → new feedback**. MCP can connect the Agent to the SaaS, but the policy, permission, verification, and recovery layers are what could make the system truly autonomous.

In conclusion

This is still just an experimental idea, not a proven way to build SaaS. The discussion also highlighted some of the hard problems: user feedback can be ambiguous, user interests don't always align with the business, and giving an Agent too much control could easily create stability and security risks.

I don't know yet how far a self-evolving SaaS can actually go. But I think it's an interesting direction worth experimenting with, especially for products that are still iterating quickly.

If you have thoughts or want to discuss the idea, feel free to follow me on X: @alwuyy.