Damien Murphy

Damien Murphy is a full-stack and applied AI engineer who builds real-time voice agents and autonomous business workflows. After working as a senior applied engineer at Deepgram, he joined Bench Computing as a founding engineer, applying lessons from conversational AI to enterprise systems that coordinate specialized agents, external services, and sensitive business data.

At Deepgram, Murphy worked directly with customers deploying speech recognition, language models, and text-to-speech in production. His voice-agent workshop demonstrates a browser-based ordering assistant that streams microphone audio over WebSockets and uses model-generated function calls to update orders. Its public client and server expose the complete architecture: browser audio capture, speech playback, order management, and an Express backend. He also collaborated on a podcast-processing automation that combined transcription, summarization, and synthesized speech.

By the 2025 AI Engineer World’s Fair, Murphy had joined Bench Computing, where his focus expanded to multi-agent business automation. His A2A and MCP workshop shows how a host agent can process meeting transcripts, delegate tasks, create GitHub issues, post Slack messages, and research companies.

• Conversational timing is a product requirement. Murphy treats interruption handling, audio bandwidth, geographic deployment, and time to first token as fundamental design constraints. His emphasis on semantic endpointing recognizes that hesitation, backchanneling, and genuine completion require different responses.

• Specialized agents prevent context overload. Murphy assigns narrow responsibilities to subordinate agents, allowing them to process verbose records or tool responses and return concise results. This context isolation for specialized agents reduces latency and limits unnecessary exposure of intermediate data while supporting parallel execution when tasks are independent.

• Protocols must justify their complexity. He distinguishes MCP, which connects models with tools and context, from A2A, which supports discovery and delegation between independently managed agents. When developers control the entire system, he favors direct function calls; protocols become useful when third-party integrations or remote agents require a shared interface.

• Cost and safety depend on operational details. Murphy evaluates prompt caching and context management against actual usage patterns, routes simpler tasks to less expensive models, and anticipates silent integration failures. His safeguards include user-scoped OAuth permissions, synthetic accounts for realistic testing, verification of external actions, and human confirmation for consequential changes.