Frequently Asked Questions
Common Questions About Building a Context Engine for AI Agents
Everything you need to know before registering.
What are the five core components of the context engine built in this workshop? +
The five core components are: (1) the Semantic Blueprint Generator that creates structured agent specifications from task descriptions, (2) the Context Router that assembles the appropriate context package for each agent invocation, (3) the MCP Orchestrator that dispatches tasks to specialised agent servers and collects typed results, (4) the Knowledge Layer comprising the RAG pipeline and memory manager, and (5) the Glass-Box Logger that records every component operation with structured metadata for observability and debugging.
How long does it take to build each component of the context engine? +
The context engine is built incrementally across the 6-hour session: the Glass-Box Logger is built first (30 minutes) as the foundation for observing everything else, the Semantic Blueprint Generator next (45 minutes), followed by the Knowledge Layer (60 minutes), the MCP Orchestrator (60 minutes), the Context Router (45 minutes), and integration testing and production preparation (the final 60 minutes). Each component is tested before the next is built.
How do the five context engine components connect to each other? +
The components connect through a pipeline architecture: a user request enters the MCP Orchestrator which invokes the Semantic Blueprint Generator to create agent specifications, passes those specifications to the Context Router which assembles context packages from the Knowledge Layer, dispatches agent invocations through MCP with the assembled context packages, and collects results. The Glass-Box Logger wraps every interaction between components, creating a complete audit trail of the entire pipeline execution.
What happens in the context engine when a specialised agent is unavailable? +
When a specialised agent is unavailable (MCP server health check failure or timeout), the context engine's fault tolerance logic activates: it checks whether a fallback agent server is configured for the same capability, routes the task to the fallback if available, records the primary failure and fallback routing in the Glass-Box log, and returns a partial result to the orchestrator if no fallback is available rather than failing the entire workflow. The workshop covers implementing this fault tolerance logic.
How do I extend the context engine with a new type of specialised agent? +
Extending the context engine with a new specialised agent involves: implementing the new agent as an MCP server with typed tools, registering the server address in the context engine's agent registry, adding a capability description to the orchestrator's planner blueprint so the orchestrator knows when to route tasks to the new agent, and adding the appropriate knowledge domain to the RAG pipeline if the new agent requires domain-specific retrieval. No changes to the core context engine components are required.
How does the context engine built in this workshop compare to commercial AI orchestration platforms? +
The context engine built in this workshop is architecturally simpler and more transparent than commercial orchestration platforms, but provides the same core capabilities for most use cases: multi-agent task decomposition, typed agent coordination, RAG-grounded knowledge retrieval, memory management, and comprehensive observability. Building it yourself gives you complete understanding of the system, full customisability for your specific use case, and no vendor lock-in or usage fees.
