AgentIR: A Compiler-Inspired Optimization Framework for AI Workflows

1. Modern Large Language Model (LLM) applications are increasingly built as complex workflow graphs using frameworks such as LangGraph, CrewAI, and the OpenAI Agents SDK. These workflows often involve multiple LLM calls, tool integrations, retrieval steps, and conditional branching. However, despite their complexity, they are typically executed exactly as defined, without any form of optimization.

1. Unlike traditional programming languages, where compilers analyze and optimize code before execution, current AI workflow frameworks lack a compile-time optimization layer. This makes it difficult to identify redundant steps, improve execution strategies, or systematically analyze performance bottlenecks ahead of runtime.

1. This talk proposes a new approach through AgentIR, an open-source framework inspired by compiler design principles. AgentIR introduces concepts such as Intermediate Representations (IR), static analysis, dependency tracking, and optimization passes into the domain of AI workflows.

1. During the session, we will demonstrate how AI workflows can be transformed into a unified IR, enabling structured analysis and optimization. Attendees will see how reusable compiler-style passes can be applied to eliminate inefficiencies, detect redundant operations, and generate more efficient execution plans.

1. The talk will include practical demonstrations such as workflow visualization, dead node elimination, duplicate tool-call detection, dependency analysis, parallel execution planning, and workflow benchmarking using open-source Python tools.

1. By the end of the session, attendees will gain a clear understanding of how compiler techniques can extend beyond traditional programming languages, how graph-based optimization can significantly improve AI workflows, and how to design compiler-inspired infrastructure for modern AI systems.

Compiler Concepts Beyond Programming Languages

Learn how Intermediate Representations (IR), static analysis and optimization passes can be applied to AI workflows.

Designing AgentIR

Understand how AI workflows can be transformed into a common Intermediate Representation for analysis and optimization.

Workflow Optimization

See how compiler-inspired techniques like dependency analysis, dead node elimination and parallel execution planning improve AI workflows.

Building Compiler Infrastructure

Explore the architecture behind a modular optimization pipeline for modern AI workflow systems.

From Theory to Implementation

Watch a complete workflow transformed, optimized, benchmarked and executed using an open-source prototype.