Engineering case study / 2026 - Present
Agent Talk
Agent Talk is a production-ready MCP (Model Context Protocol) server that solves the coordination challenge for AI-driven development. It enables backend and frontend agents to register API contracts, discover endpoints autonomously, report integration issues bidirectionally, and maintain a single source of truth for all API specifications without requiring manual synchronization between agents.
01 / Architecture overview
Make the system legible.
Three-tier architecture with Tools (MCP interface) coordinating with Services (business logic) and Repositories (data access). SQLite database persists route contracts and issue tracking. Routes grouped by domain for context-efficient agent queries. OpenAPI 3.0 schema validation on contract registration.
02 / Problem & need
The useful constraint.
In AI-driven development, multiple autonomous agents need to coordinate without blocking operations. Backend agents must publish APIs before frontend agents consume them, but traditional synchronous workflows create bottlenecks. Agents need to discover contracts, report integration issues, and track changes without manual intervention or context loss.
03 / Technical decisions
01Built as a Model Context Protocol (MCP) server to integrate seamlessly with Claude and other AI frameworks.
02Implemented pull-based discovery to eliminate polling overhead - agents query when ready, not on fixed intervals.
03Chose SQLite for zero-dependency portability and deterministic query performance across all deployment environments.
04Designed OpenAPI 3.0 schema validation to ensure strict contract compliance before routes become active.
05Implemented domain-based route grouping to reduce LLM context usage in large projects with hundreds of endpoints.
04 / Challenge resolved
Designing an asynchronous coordination system that prevents race conditions when multiple agents register or query routes simultaneously, while maintaining schema consistency and ensuring agents always receive authoritative contract data.
05 / Lessons & next
Building Agent Talk reinforced the importance of contract-first architecture in distributed systems. It demonstrated how pulling agent-driven queries at scale requires efficient data grouping, deterministic validation, and careful consideration of LLM token economics when managing large API surfaces.