Collaborative Design with Autonomous Agents

PhD Dissertation, Stanford University, 1997
Department of Mechanical Engineering. Advisor: Mark Cutkosky.

A human designer can be greatly assisted in the intricate task of routing cable harnesses through complex mechanical assemblies if empowered to interact naturally with other autonomous agents. This dissertation presents a system in which specialized computational agents collaborate with a human designer to solve problems too complex for any single approach, whether human or algorithmic.

3D cable harness simulation showing dense bundles of routed cables with green connectors — The cable harness routing problem: finding optimal paths for bundles of cables through complex mechanical assemblies.

The Architecture

Figure 1.8 from thesis: Creators, Mutators, and Combinors feeding into a centralized Design Pool containing Graph Routing Representations — Figure 1.8 -- Multiple agents can build and reuse designs stored in a centralized pool.

Three types of agents handle different aspects of the problem:

Creators generate new solutions from scratch. "Go build this thing."
Mutators make small improvements to existing solutions. "Fix this bug." A small change on a big thing.
Combinors restructure and merge the best parts of different solutions. "Restructure this code."

The human designer is not outside the system. The human is part of it. Agents handle the computationally intensive search and optimization, while the human contributes qualitative judgments that agents cannot model: spatial intuition, aesthetic preferences, manufacturing knowledge, and the ability to recognize when a solution "looks right" even before formal metrics confirm it.

The design pool serves as shared memory. All agents, including the human, deposit solutions into the pool and draw from it. Each agent builds on the work of others without requiring direct coordination between agents.

Why this matters now (2026): This is exactly how AI-assisted coding works today. When Andrew rebuilt FriendFinder in 25 days using AI tools, the process was the same architecture from his 1997 thesis. He acts as the human in the loop -- providing vision, taste, and judgment. The AI acts as Creator ("build this page"), Mutator ("fix this bug"), and Combinor ("restructure this component"). The system gets better because the human thinks of things the AI isn't thinking to do. Twenty-seven years later, the same principles, at a different scale.

Key Finding

Cable bundle routing visualization showing optimized paths through a complex assembly — Optimized cable routing -- the result of human-agent collaboration producing better solutions than either could achieve alone.

Experimental results demonstrated that a human working collaboratively with specialized agents produced better solutions than either the human or the agents working alone. The collaborative system consistently found higher-quality cable routings in less time than purely manual design or purely automated optimization.

Three Principles

Decomposition. Complex problems become tractable when broken into specialized subtasks handled by purpose-built agents.
Collaboration. Multiple approaches working through a shared pool outperform any single method, including the human designer alone.
Evolution. Solutions improve through iterative selection and recombination of successful patterns across the design pool.

The question of how humans and autonomous agents collaborate effectively remains as relevant today as it was in 1997. The through-line is always the same: how do you get separate entities to find each other and build something better together?