Over the last two weeks, I set out to evolve MyNewsAI.io — a Generative AI-powered news platform — into something more intelligent and autonomous. What started with content summarization and translation began to scale into what many are calling Agentic AI: systems that don’t just generate — they act.
But moving from GenAI to Agentic AI wasn’t just a feature upgrade. It exposed deep technical, architectural, and strategic hurdles that most devs and AI founders will likely face too. This post is my reflection — and perhaps, a roadmap for those on the same path.
🔧 Technical Hurdles
1. Inference Speed and Scalability
Even after moving from CPU to GPU instances on AWS (e.g., g4dn.xlarge), inference remained slow.
Tasks like Wav2Lip video rendering, multi-language translation, and model chaining are compute-heavy, and real-time performance remains a challenge — both in speed and cost.
2. Incomplete and Fragmented Documentation
Documentation across the GenAI ecosystem is sporadic. From installing CUDA dependencies to chaining models, I found myself relying heavily on ChatGPT, but it doesn’t always know the latest packages or AWS configurations.
3. Lack of Robust Open Source Agentic Tools
There’s no “Apache for Agentic AI” — yet. LangChain, CrewAI, and similar tools are exciting but feel early-stage. Production-grade agent systems need mature orchestration, memory, retry logic, and error handling — and we’re not quite there.
Architectural Challenges
1. LLMs Aren’t Agentic by Nature
Large Language Models like GPT, Claude, or LLaMA are generators, not agents. They can write summaries or responses, but they struggle to plan, reason, or act autonomously unless given very careful prompt scaffolding and external control.
2. Paywalled Content: Data Access Is Broken
Most quality news sources now sit behind paywalls. For a news AI agent, that limits reliable input. We need solutions that respect content rights but still allow agents to access trusted, real-time knowledge.
Strategic Limitations
1. We Need Open Agent Stacks
Just as Apache catalyzed the Java ecosystem, Agentic AI needs a modular, open-source foundation: shared protocols, middleware, and community support. Today, every AI dev is reinventing the same scripts.
2. LLMs Need a “T Model” Mindset
LLMs today are often deep in one topic but lack breadth across domains. True Agentic AI needs T-shaped intelligence: horizontal coverage across multiple domains with vertical depth where needed — especially in news, where politics, finance, and culture intersect.
🚀 What's Next for MyNewsAI.io?
Moving toward Agentic AI isn’t just about upgrading the model — it’s about rethinking the system.
At MyNewsAI.io, I’m building toward:
- Multi-step agents (summarize → translate → synthesize → narrate)
- A real-time anchor that can present news via lip-synced video
- Eventually, a fully autonomous AI newsroom assistant
The work is hard, but the potential is massive.
Join the Journey
If you're working on Agentic AI, multi-modal pipelines, or open-source infrastructure — let’s connect.
You can follow the project at https://mynewsai.io, or ping me if you want to collaborate.
