Making AI Deterministic for Developers and their Agents, with Patrick Vuong of Moderne

Episode Summary

Executive Summary: The episode centers on Modern’s shift from code modernization to agent tooling, with Director of Product Patrick Wong explaining how the company helps AI agents work more accurately and efficiently at scale. He argues that as software development becomes more agent-driven, enterprises need deterministic tools, semantic code context, and orchestration layers to make AI outputs safe, auditable, and repeatable.

Main Topics: Patrick Wong’s career path and product mindset (Priority: 4/5): Patrick shares his background in Toronto, his business education at McMaster, and an eight-year Microsoft career spanning marketing, go-to-market, product marketing, and developer tooling. A recurring theme is his focus on understanding customer needs and turning them into product insight. Modern’s origin in code modernization (Priority: 5/5): Modern began by helping teams reduce tech debt and modernize codebases using OpenRewrite, especially for large-scale framework upgrades, vulnerability fixes, and long-term maintainability. This foundation is positioned as a prerequisite for AI readiness. Why AI agents need tooling (Priority: 5/5): Patrick explains that agents need tools because inference alone is non-deterministic, expensive in tokens, and weak at large-scale context retention. Modern’s thesis is that agents need accurate, efficient, deterministic capabilities to operate reliably in enterprise environments. Deterministic semantic model and recipes (Priority: 5/5): Modern’s core technical approach is built on a lossless semantic tree, described as a compiler-accurate model of code with full type resolution and cross-file understanding. On top of that, Modern creates 'recipes'—programmable deterministic transformations rather than prompts. Product suite for agent effectiveness (Priority: 4/5): Modern’s current offerings include Trigrep for symbol-aware code search, Prethink for precomputed architectural context, and ChangeLog for portfolio-level awareness. These tools are delivered through MCP to make coding agents more effective and less wasteful. Future vision: agent factories and governance (Priority: 5/5): Modern’s roadmap includes deeper deterministic tooling, multi-repo execution, production vulnerability remediation, and 'modern factories' where agents and humans collaborate to modernize systems autonomously. It also emphasizes agentic SDLC artifacts like transcripts, change logs, and audit trails. Career advice and category creation (Priority: 3/5): Patrick closes by emphasizing that product teams should build categories, not just ship isolated features. In fast-moving AI markets, every feature should reinforce the larger product story and strengthen market understanding.

Key Arguments: Agents need tooling because raw model inference is non-deterministic and expensive, making accuracy and efficiency critical at scale. Legacy or messy codebases reduce AI agent effectiveness, so modernization is a prerequisite for successful agentic development. A lossless semantic tree gives agents compiler-grade understanding of code, enabling reusable context at enterprise scale. Recipes are superior to prompts for certain tasks because they are deterministic programs that can safely drive repeatable transformations. The SDLC was built for humans, so agent-driven development needs a new harness that preserves governance while supporting speed and volume. Modern’s products help agents search better, understand architecture faster, and avoid duplicating work across a portfolio. The future of software operations will blend agents and humans in orchestrated factories, with institutional memory becoming a durable competitive advantage.

Data Points: Microsoft tenure: 8 years - Patrick Wong described spending eight years at Microsoft across multiple roles and regions. Customer insight program cohort: 12 students - At McMaster, Patrick joined an experiential learning program that selected 12 students and split them into two teams. Program structure: 2 teams - The experiential learning program divided students into two teams to solve weekly company business cases. AI transformation adoption rate: around 8% - Patrick cited MIT’s reported figure that large-scale AI transformation is at about 8%. Prethink reasoning improvement: 4x faster reasoning - Patrick said Modern’s Prethink product has produced four times faster reasoning. Prethink tool-call reduction: 55% fewer tool calls - Patrick cited this reduction as a result of Prethink’s precomputed architectural context. Prethink token reduction: 30% fewer tokens used - Patrick said Prethink reduces token usage by 30%.

Pivotal Quotes: "These are programs, they're not prompts." — Patrick Wong: He contrasted Modern’s recipe-based approach with traditional prompt engineering when explaining deterministic code transformations. "The harness makes sure the work actually lands." — Patrick Wong: He described Modern’s role in agentic SDLC as a deterministic harness that ensures agent-produced work is safe and usable. "What is that story you're driving?" — Patrick Wong: In his advice to others entering the field, he emphasized building a category narrative rather than marketing isolated features.

Implications: The episode suggests AI software development is shifting from model-centric demos to infrastructure that makes agent behavior reliable, auditable, and scalable. For teams, the message is clear: modernize codebases, add deterministic tooling, and design for an agentic SDLC.