Agenda
Redesigning Legacy APIs for Humans, Scripts, and Agents
Bernard Rodrigue (Director, Wwise Experience, Audiokinetic Inc.)
Pass Type:
Academic Pass, Festival Pass, Game Changer Pass, Indie Pass, Start-Up Pass -
Get your pass now!
Track:
Audio
Audience Level:
Advanced
Format:
Lecture
Vault Recording: TBD
Audience Level: Advanced
With agentic workflows becoming a reality in game studios, developers are looking to expose their engines and tools to Model Context Protocol (MCP) wrappers. However, exposing massive, complex, legacy codebases to LLMs often leads to high token usage, slow performance, and hallucinations.
When we attempted to expose our own editor API to an LLM, we discovered that AI agents fail at using our tools for the exact same reasons human developers do: the information is fragmented, difficult to find, and the API is too complex.
In this session, attendees will learn concrete techniques to retrofit proprietary engines and tools for AI agents. Sharing specific failures and pivots, we will detail how we optimized token usage via dynamic tool discovery, handled massive projects using sandboxed Lua execution, and benchmarked our MCP within our CI test environment. Ultimately, optimizing an API for agents fundamentally improves the human developer experience.
When we attempted to expose our own editor API to an LLM, we discovered that AI agents fail at using our tools for the exact same reasons human developers do: the information is fragmented, difficult to find, and the API is too complex.
In this session, attendees will learn concrete techniques to retrofit proprietary engines and tools for AI agents. Sharing specific failures and pivots, we will detail how we optimized token usage via dynamic tool discovery, handled massive projects using sandboxed Lua execution, and benchmarked our MCP within our CI test environment. Ultimately, optimizing an API for agents fundamentally improves the human developer experience.
Takeaway
Attendees learn concrete strategies to retrofit complex legacy APIs for AI agents. Developers discover how to minimize token consumption via dynamic tool discovery, prevent domain hallucinations by injecting proprietary data models, process massive asset counts safely using Lua sandboxing, and benchmark non-deterministic models within standard deterministic CI/CD pipelines.
Intended Audience
This session targets tools programmers, software developers, and technical sound designers integrating Agentic AI into studio pipelines. Attendees gain actionable blueprints for retrofitting massive legacy APIs for LLMs. Prerequisite knowledge includes familiarity with general API architecture, scripting concepts, and basic AI workflows to grasp the technical content.