OpenAI’s recent push toward increasingly capable AI agents is raising an interesting question about the future of software.

Traditionally, users interact directly with SaaS platforms, CRM tools, analytics dashboards, project management software, and productivity suites. But if AI agents can access multiple systems, retrieve information, execute tasks, and coordinate workflows on a user’s behalf, do we still need to interact with those applications directly?

Some industry leaders are suggesting that the future interface may not be software itself, but an AI layer that sits on top of software.

That could fundamentally change how businesses think about:

• Enterprise applications
• User interfaces
• Data access
• Workflow automation
• Software purchasing decisions

Will AI agents become the primary way people interact with business systems, or will SaaS platforms remain the center of enterprise work with AI acting as an enhancement layer?

Curious to hear where others think OpenAI’s agent-driven approach could take the industry over the next few years.

As machine learning pipelines become more complex, I find myself struggling with a design question rather than a coding one.

When you’re working with scikit-learn’s Pipeline and ColumnTransformer, where do you place custom feature creation logic?

For example:

• Creating interaction features
• Extracting date-based features
• Combining multiple columns into a new feature
• Domain-specific transformations

Some practitioners add these steps before the ColumnTransformer, while others treat feature engineering as part of preprocessing and keep everything inside a single pipeline.

On one hand, keeping everything in the pipeline improves reproducibility and prevents training-serving skew. On the other hand, deeply nested transformers can become difficult to debug and maintain.

I’m curious how experienced ML engineers structure their workflows:

• Do you separate feature engineering from preprocessing?
• Do you use custom transformers extensively?
• How do you keep pipelines both reproducible and understandable as projects grow?

Interested in hearing real-world approaches, especially from teams managing large production ML workflows.