The Problem
As AlgoEd scaled to serve thousands of schools globally, internal operations became a bottleneck. Setting up data on the internal admin panel involved repetitive, multi-step API calls — creating records, linking entities, populating fields in a specific sequence. Two core issues made this unsustainable:
- Manual and error-prone. Each setup required careful manual work across multi-step API sequences, with mistakes requiring full restarts.
- Too variable for scripts. Setup sequences varied by context, making a single hard-coded script impractical — every variation needed its own logic.
- Built-in obsolescence. AI technology evolves monthly. Any rigid, bespoke tooling risked obsolescence within months if tightly coupled to a specific model or framework.
Design Philosophy
Rather than building rigid automation, we adopted a principle of dynamic flexibility — structure everything to be modular and swappable, right-size agent complexity to task complexity, and design for replaceability over permanence.
Dynamic Flexibility
Modular, swappable components. Avoid locking into any single model or framework so the system evolves with the technology.
Right-Sized Agents
Match agent complexity to task complexity. Manus for long-running multi-step tasks, Claude for simpler well-scoped operations.
Don't Over-Engineer
Ship working solutions fast. Accept that the best tool today may be replaced tomorrow — design for that reality.
Security by Design
Expose only what's necessary. Anonymize by default. Layer access controls at the application level.
The Solution
We built a four-layer architecture that lets an AI agent autonomously execute complex admin workflows:
Designed API Pathways
Identified and documented the exact API endpoints needed for each setup workflow. These pathways were purpose-built to be consumed by an AI agent, not a human operator.
Built an MCP Server
Wrapped the API pathways in an MCP (Model Context Protocol) server — a structured interface that any compatible agent can call. The MCP layer handles authentication, request formatting, and response parsing.
Created Agent Skills
Wrote detailed skills documents giving Manus the context it needs — what each endpoint does, required sequencing, edge cases, and validation rules. Skills serve as modular, updatable knowledge bases.
Deployed Manus as Executor
Manus, armed with skills docs and MCP access, autonomously executes multi-step data setup sequences — handling branching logic, retries, and validation without human intervention.
The Results
The architecture transformed how AlgoEd handles internal operations:
Key Takeaways
- MCP as the universal interface layer. By standardizing on MCP, we can swap agents, models, and frameworks without rebuilding integrations. The MCP server is the stable surface; everything above it can change.
- Agent Skills as the knowledge layer. Instead of hard-coding business logic into prompts, skills documents serve as modular, updatable knowledge bases that any agent can consume.
- Right-size the agent to the task. Manus for complex multi-step workflows, Claude for simpler operations — avoiding unnecessary cost and latency.
- Build for replaceability, not permanence. In a landscape where models evolve monthly, the most durable architecture is one designed to be easily modified.
Ready to see similar results?
Book a free consultation and discover how SnapTask can transform your operations — or get your money back.