Agentic AI disruption of the dev workflow

As a lifelong programmer, I find myself living in two worlds at once: one where AI has become an essential co-pilot in my daily work, and another where its limitations still remind me who’s actually steering the ship. Every day I see firsthand how machine learning models and coding assistants can accelerate development, generate insights, and even spark creativity - yet just as often, I see how hallucinations, privacy concerns, or opaque decision-making can derail a project faster than a bad merge.

I recently sat down with Sam Basu, Principal Developer Advocate at Progress Software, to discuss how developers can thoughtfully integrate AI into their applications, from managing model dependencies and monitoring production systems, to understanding the rise of agentic AI and its impact on developer roles.

Whether you’re building for mobile, web, or enterprise, the conversation is a practical roadmap for navigating the intersection of innovation and responsibility in AI-driven development - where every prompt, every deployment, and every decision still needs a human in the loop.

ADM: As agentic AI and It's disruption of the dev workflow become increasingly capable, what practical considerations should app developers keep in mind when integrating them into production environments? 

Basu: AI Models can be treated like a service dependency - monitored for metrics & performance. Production integration/deployments need to consider firewall, privacy & ethical concerns such as bias or hallucination risks. Developers should also plan for versioning and fallback logic in case the AI service degrades or changes unexpectedly.

ADM: What are some common deployment pitfalls specific to mobile or web app contexts? 

Basu: Small Local Models are nifty to deploy with mobile/web apps, but need close monitoring for memory/storage/CPU/GPU spikes. Resource-constrained environments make performance tuning crucial, and developers should also ensure fallback behavior if models fail or produce poor results.

ADM: How can dev teams effectively test and monitor AI features post-launch? 

Basu: Post-deployment monitoring of AI services should be similar to other dependencies, such as Kubernetes and .NET. The Aspire dashboard can help. In addition, teams should perform continuous validation against updated datasets and ensure compliance with evolving privacy, security and data policies.

ADM: With the rise of agentic AI systems - tools that can reason, plan, and act - how do you see developer workflows evolving? 

Basu: AI Agents can have a significant productivity boost for developer workflows - tasks such as getting started, routine integrations and repetitive tasks can be delegated with supervision. This allows developers to spend more time on innovation and problem-solving.

ADM: Could these agents replace parts of traditional CI/CD pipelines? 

Basu: Agentic DevOps is more about inserting AI Agents for autonomous actions throughout the pipeline, while still requiring human intervention. We may be a bit away from complete CI/CD automation, but agents can already assist with testing and build optimization.  

ADM: How might junior vs. senior dev roles shift in response to these tools? 

Basu: AI Agents can do much of the mundane, freeing humans to do complex, creative and coordination tasks. Teams can start thinking of AI Agents as a very junior developer on the team. They are capable of handling well-scoped tasks but still need guidance and context from experienced developers.

Feeding the Machine: Solving Contextual Challenges in AI Integration 

ADM: One of the trickiest parts of working with LLMs is providing the right context. What are the top challenges developers face here, and how can they be overcome?

Basu: LLM knowledge is often timestamped and lacks local context. Instruction and setup files can help. AI Agents can use contextual tools through MCP Servers. Developers also need to inject relevant domain-specific data proactively at runtime.

ADM: Are there effective architectural patterns (e.g., RAG, vector stores) you recommend? 

Basu: Vector DB searches are often highly effective. However, RAG solutions require careful balancing when scaling to large workloads. For narrower or well-defined scopes, RAG architectures can be particularly effective.

ADM: How do you approach managing session memory and user history context? 

Basu: Generic AI Models or services on top of them provide huge context windows these days - for custom implementations, session context comes down to the cost of memory and tokenization.

ADM: With so many AI coding assistants available—from GitHub Copilot to AWS CodeWhisperer—how should developers evaluate which one is best for their stack or workflow? 

Basu: No shortcuts - developers need to make sure coding assistants are compliant, check the value they bring and test them in real-world scenarios. Performance and reliability under pressure are key factors, and what works for one team may not suit another.

ADM: Should teams choose tools based on IDE compatibility, model capability, or privacy guarantees? 

Basu: Privacy is essential, and AI Models have unique strengths - some trial and error helps. Developers should be able to use whichever IDE/tooling that adds productivity without changing workflows, but they also need to make sure the model’s behavior aligns with company policies.

ADM: Do you see a future where developers regularly use multiple AI assistants in tandem? 

Basu: Not really - at the end of the day, developers have to deliver software on time. Most will evaluate and settle on one for productivity, as switching between tools can become counterproductive unless the assistants are tightly integrated into one workflow.

ADM: Off-the-shelf AI tools can only go so far. What are some ways developers can fine-tune or extend these tools to match their team’s unique needs and coding standards? 

Basu: Custom AI models can be helpful, but expensive. Instruction files are a great way to add context and coding guidelines for AI. Developers will choose tools that allow extensibility and easy integration.

ADM: Have you seen effective use of custom prompt libraries, fine-tuning, or plug-in APIs? 

Basu: RAG solutions can be hard to scale. Telerik & KendoUI AI Coding Assistants are effective GitHub Copilot Extensions or MCP Tools that bring custom knowledge & context to the table.

ADM: What role does team culture play in AI adoption and standardization? 

Basu: Cultural buy-in is key. Teams must be open-minded to adopt AI – it's augmenting productivity, and mileage will vary. With the right customizations, AI coding assistance gets better over time.

ADM: AI promises faster development, fewer bugs, and more efficient code. In your experience, how much of this promise holds up in real developer workflows? 

Basu: Real-world AI often has the 90-10 rule. It is easy to get 90% of things right; the last 10% is more challenging. Still, AI is increasingly helping developers by handling the routine, repetitive work and giving them a stronger starting point for their projects.

ADM: Are there specific types of tasks where AI consistently under- or over-performs? 

Basu: Getting new projects started with specific frameworks and dependencies is something AI has gotten very good at. Underperformance is often tied to deeply contextual questions with existing code. Understanding application-specific logic or navigating legacy systems remains challenging for most models.

ADM: How are teams measuring ROI when integrating AI into their dev process? 

Basu: Engineering teams are starting to treat Agentic AI as a team member or part of DevOps. ROI can be measured in terms of the expectations that a team would have for a newbie developer.

ADM: From GPT-4 and Claude to open-source models like Code Llama and Phi-3, there’s no shortage of options. Which AI models stand out for developer use cases, and how do they differ?

Basu: AI Models are not created equal and the best fit depends on use cases. GPT/Gemini has proved to be effective for code generation, while Claude seems to do better for long reasoning tasks.  

ADM: Are smaller, local models becoming viable for production use? 

Basu: Machine Learning is still viable & companies may decide to use local AI Models for specialized training/privacy reasons. Models are getting nimble, but general LLMs still have the most flexibility.

ADM: What’s your take on proprietary vs. open models when it comes to control and reliability?

Basu: Proprietary models offer customizations, privacy & control - good for unique use cases. The public AI models are more powerful in NLP and general-purpose usage.