Hi everyone,
I’ve been working on a personal project over the past few months and would love to get feedback from the community.
I’m a data and automation specialist working in enterprise analytics, and I’ve been experimenting with a different approach to agentic AI systems.
GitHub (MVP version):
Most AI solutions today seem to be built around a single LLM, where the model effectively becomes the centre of the entire solution. While that works well for many use cases, I kept running into the same questions:
- What happens when a better model becomes available?
- What if a cheaper model can do 80% of the work?
- What if certain workloads require higher confidentiality?
- What if different models have different strengths?
- What if the real value isn’t the model, but the workflow surrounding it?
This led me to build an MVP focused on a simple principle:
The workflow should own the intelligence, and the LLM should be a replaceable component.
Instead of:
User
↓
Single LLM
↓
Output
I explored an architecture more like:
User
↓
Agent Orchestration
↓
Memory & Context
↓
Governance Layer
↓
Tool Execution Layer
↓
One or More LLMs
The key idea is that business processes, governance, memory, and orchestration remain stable even while models change underneath.
Why I Think This Matters
The AI industry currently spends a lot of time comparing models.
But models are improving so quickly that today’s leader may not be tomorrow’s leader.
If organisations tightly couple their workflows to a single provider, switching later can become expensive and disruptive.
A decoupled architecture potentially enables:
- Multi-model orchestration
- Cost optimisation
- Reduced vendor lock-in
- Easier model replacement
- More controlled governance
- Better enterprise integration
For example:
Planning Agent → Model A
Coding Agent → Model B
Reviewer Agent → Model C
Documentation Agent → Model D
Each agent can use the model that offers the best balance of quality, speed, cost, or privacy.
An Interesting Side Effect: Confidentiality
One aspect I don’t see discussed often is information segmentation.
In a traditional architecture, a single provider may receive:
- Requirements
- Business rules
- Internal documentation
- Generated outputs
Assembling the full picture.
In a distributed agent architecture, different pieces of work can potentially be routed to different models or providers.
For example:
Requirements Analysis → Model A
Code Generation → Model B
Validation → Model C
Documentation → Model D
No individual provider necessarily sees the complete business context.
Obviously, this doesn’t replace proper security controls, governance, or legal agreements.
However, it may provide an additional layer of confidentiality by reducing how much information is exposed to any single model vendor.
I’m curious whether anyone else has experimented with this approach and whether you’ve seen practical benefits.
MVP vs Full Edition
The public GitHub repository is intentionally an MVP.
It demonstrates concepts such as:
- Multi-agent orchestration
- Context management
- Tool execution
- Memory handling
- Workflow design
- Extensible architecture
Internally, I’ve also developed a more comprehensive edition focused on enterprise use cases, including:
- Human-in-the-loop approvals
- Governance controls
- Audit logging
- Role-based access control
- Organisational memory
- Cost and usage monitoring
- Agent lifecycle management
- Workflow analytics and dashboards
- Enterprise system integrations
The goal isn’t to build another chatbot.
The goal is to build a reusable operating framework for AI-assisted business workflows.
Problems I Believe This Could Help Solve
- Organisations becoming dependent on a specific model vendor
- High AI operating costs
- Difficulty governing AI actions
- Poor visibility into AI decision making
- Lack of auditability
- Rebuilding workflows every time a new model appears
- Challenges scaling AI from experimentation into production
Questions for the Community
I’d love to hear your thoughts on a few questions:
- Do you think workflow orchestration will become more valuable than individual model selection over time?
- Have you implemented multi-model architectures in production?
- How are you balancing cost, performance, and governance?
- Are there enterprise challenges that you believe current agent frameworks still don’t address well?
- Do you see confidentiality benefits in distributing workflows across multiple models, or do you think the added complexity isn’t worth it?
Looking forward to hearing different perspectives and learning from others who are building in this space. Thanks for reading.