CMI Consulting Group
In a May blog post from Elasticsearch, Matthew Skinner highlighted the major advantages of using Elasticsearch as the backend platform for an Enterprise AI stack.
https://www.elastic.co/blog/elasticsearch-platform-missing-piece-ai-stack
It highlighted the four main memory types an AI stack uses and how ELK stack fits in as one and single engine for all of them.
Episodic memory
Semantic memory
Procedural memory
Workflow state
Consolidating all four components into one system has clear advantage of reducing cost, overhead and simplify. But what will be the price you have to pay for that? What are the impacts when different workload competes on the same platform?
I asked Claude to give us a deeper analysis on the pros and cons for using Elasticsearch as the backend platform and the answers are revealing. It is a good complement for the Elastic blog post.
As Agentic AI transforms from experimental sandbox into production-grade operational capability, building a solid enterprise architecture for this ecosystem becomes both critical and urgent. I asked Claude to research and synthesize guidance from Gartner, LangChain, Microsoft, and Anthropic into a practical reference architecture for internal enterprise AI systems.
The architecture addresses four questions:
What are the main components and layers of the architecture?
What are their functions and roles?
What are the best practice design considerations for each capability?
What are the key governance-related capabilities?
Each layer is covered in depth, with design recommendations and a readiness assessment for enterprise architects.
With Agentic AI, programming and debugging enter a new era where machine takes over main part of the heavy lifting work. For those that have tested Vibe Coding, you will notice the new challenge this ways of working creates. It is about debugging and troubleshooting, or in general observability of the code execution. Especially if you start to incorporate more and more of 3rd party Skills and MCP services.
Again, I asked the Claude to share some of the latest research within this area, as well as the best practice recommendations.
After digging in some of the new released observability AI agent solutions for some days, one question popped up in my mind, is Agentic AI really production ready? We know seeing an application function in a demo is one thing, but making it to run 24x7 for mission critical tasks in enterprise is something completely different. My personal app built by Claude Code suffers as well from intermittent reliability problems even though I have incorporated logging, tracing and observability integration.
I posted the question to Claude, and find the answer it provides quite comprehensive. Thus I would like to share this summary here.
The newly published "The Founders playbook - Building an AI-native startup" is a very interesting read for technology startups. It not only provides clear business insight over the common early organic growth lifecycle stages of the start-ups with its typical challenges and characteristic, it provides also a lot of tips and tricks of how AI tooling is impacting the current model.
It is clear that AI tools have different impacts for organizations at different stages of its lifecycle. My key learning after going through this playbook is the deeper understanding of why the various roles and job positions are needed as organizations grow and mature. The jobs were created because of the need and demand of both customers and external stakeholders as well as regulational governance. AI will help increase efficiency and productivity, but the accountability will always be on person, and it cannot be the founders all the time.
The journey of building a successful company is about getting rid of founder dependency, and the end game will never be fully AI dependency either. Then it cannot be a successful company for investors either.
Anthropic has just released ten ready-to-run AI agent templates purpose-built for financial services.
https://www.anthropic.com/news/finance-agents
Delivered as plugins for Claude Cowork and Claude Code, these agents target the most labour-intensive workflows in the industry:
client meeting preparation
market research
financial model construction
month-end close
statement auditing
and more.
Each template ships with its own skills, connectors, and subagents — a reference architecture that firms can adapt to their own risk policies and approval flows.
The promise is striking: work that previously took months can now be completed in days.
What makes this moment particularly striking is its timing. Less than a month ago, Anthropic's Mythos model preview sent ripples of concern through the global IT security community.
Now, that same forward momentum is arriving at the doorstep of financial services.
What’s next, another industry vertical? What will be the consequence of these standardizations and AI proliferation?
In workflow digitalization or automation, the main communication methodology is API calls for data exchange with the agreed format, even internally between the software. To get the outcome with the continuously datafeed and reasoning interactions, MCP (Model Context Protocol) became the new standard. Delivering intelligent AI based services with domain expert knowledge through MCP and charge for them become an interesting business model.
I asked Claude to help me summarize some of the pioneers within this field with their business and pricing model for inspiration.
Just as SAFE finally established as the new standard for ways of working and planning in IT industry, arrival of Agentic AI is shaking some of the fundaments of its manifesto and assumptions.
The whole SAFE framework was built upon the assumption that bottleneck for software development is upon the developer and engineering resources in planning, collaboration, productivity and availability. To cope with this, SCRUM and SAFE established ceremonies daily scrum, backlog refinements, PI Planning and sprint planning etc. When development cycles are weeks and months, those overhead hours spent seems to be small. But if the Agentic driven development and code generation reduces to minutes and hours, all those overheads and latency in planning and collaboration became contra productive.
To better understand the fundemental changes we are seeing on the software market (especially SaaS market side) with Agentic AI, I recommend a blog post from Julien Bek
This blog will provide answers for
Why AI-agent native software company that provides end-to-end services for different domains will triumph over the license-selling SaaS tooling firms?
A summary of Agentic AI functionalities for the major observability vendors by Claude