Design Agents

AI That Explores Design Spaces

A design agent is an AI system that explores design spaces on behalf of an engineer — generating alternatives, evaluating candidates, filtering by constraints, and presenting a curated shortlist. The agent combines the generative and optimization capabilities from DGTLENG 301 (generative design, surrogate modeling, reinforcement learning) into a coherent system that can operate with varying degrees of autonomy.

The key word is "on behalf of." A design agent is not a replacement for the engineer. It is a system that extends the engineer's reach into parts of the design space that are too large, too complex, or too tedious to explore manually. The engineer sets the objectives, defines the constraints, evaluates the shortlist, and makes the final decision. The agent does the exploration and initial filtering.

What makes a design agent different from an optimization algorithm? An optimization algorithm solves the problem it is given. A design agent reasons about how to solve the problem — it selects which tools to use, determines the exploration strategy, adapts when initial results suggest a different approach, and manages the computational budget. The optimizer is a tool the agent uses; the agent is the orchestrator.

Four Levels of Design Agent Autonomy

Design agents operate at different levels of autonomy, and the right level depends on the problem, the available tools, and the organization's trust in the agent's capabilities.

Human-Driven Design

The engineer makes all decisions. They select the design parameters to vary, choose the analysis method, define the evaluation criteria, generate candidates manually or through parametric sweeps, and evaluate each candidate. Tools assist with computation (running simulations, plotting results) but do not make decisions.

This is the baseline — traditional engineering design. The engineer's expertise is the primary asset and the primary bottleneck. The design space explored is bounded by the engineer's time, experience, and cognitive bandwidth. For simple problems with few parameters and clear trade-offs, human-driven design is efficient and appropriate.

AI-Assisted Design

The engineer drives the process, but AI handles specific subtasks. A surrogate model replaces full simulation for screening. A sensitivity analysis tool identifies which parameters matter most. A clustering algorithm groups candidates to help the engineer understand the design space structure. The engineer still selects the approach, interprets the results, and decides what to do next — but AI accelerates the individual steps.

At this level, the engineer must understand what the AI tools do and how to interpret their outputs. A surrogate prediction is not the same as a simulation result — it carries approximation error and a validity domain. A clustering result requires engineering interpretation to be meaningful. The engineer uses AI as an amplifier, not a substitute for judgment.

AI-Led Design Exploration

The agent manages the exploration process. Given a design problem (objectives, constraints, parameter ranges), the agent determines the exploration strategy: how to sample the design space, which surrogate to train, which optimization algorithm to use, when to switch from broad exploration to focused exploitation. The agent runs the analyses, filters the results, and presents a ranked shortlist to the engineer.

The engineer's role shifts to problem definition and result evaluation. The engineer specifies what the agent should optimize, reviews the shortlist, probes why specific designs were ranked highly, and selects the final design. The engineer also monitors the agent's process — checking that the exploration was sufficiently broad, that the constraints were respected, and that the agent did not get stuck in a local optimum.

This level requires trust in the agent's exploration strategy and transparency in its process. The engineer must be able to ask "why did you recommend this design?" and get a meaningful answer — not just "it scored highest" but "it achieves high stiffness because the material distribution concentrates load paths here, and it meets the thermal constraint with 12% margin because the geometry provides sufficient convective surface area."

Fully Agentic Design Exploration

The agent operates with minimal human intervention. It not only explores the design space and presents results but also adapts the problem formulation when it discovers that the constraints are contradictory, proposes alternative approaches when the initial strategy is not converging, and manages a portfolio of concurrent explorations with different strategies.

The engineer interacts with the agent at a strategic level — adjusting priorities, approving constraint modifications, and making trade-off decisions that require program context or stakeholder input. The agent handles everything else.

This level is largely aspirational in engineering today. Fully agentic design exploration requires robust reasoning about engineering trade-offs, reliable adaptation under uncertainty, and the ability to explain its decisions to engineers who must ultimately accept or reject the results. Research systems demonstrate elements of this capability; production deployment at this level is rare.

What the Agent Does vs. What the Engineer Does

At each autonomy level, the division of labor between agent and engineer shifts.

Problem formulation always starts with the engineer. The agent cannot determine what the system should do, what trade-offs the organization will accept, or what constraints are non-negotiable. At lower autonomy levels, the engineer also specifies how to explore (which tools, which methods). At higher levels, the agent determines the how while the engineer defines the what.

Constraint management becomes shared at higher autonomy levels. The engineer defines the constraints; the agent identifies when constraints conflict, which constraints are active (binding on the optimal design), and which constraints have significant margin. The agent may propose relaxing a constraint and report the performance improvement that would result — but the decision to relax is the engineer's.

Uncertainty awareness is critical for the agent. At every level, the agent must communicate not just what it recommends but how confident it is. A design that is optimal according to a surrogate with high uncertainty in that region is a different recommendation than one confirmed by full-fidelity simulation. The agent's communication of uncertainty is what enables the engineer to calibrate trust.

Design rationale changes character. In human-driven design, the rationale is the engineer's thought process. In agent-driven design, the rationale is the agent's search history, the constraints that shaped the result, and the trade-offs between competing objectives. Capturing this rationale in the digital thread (DGTLENG 103) requires new data structures — the trace of an agent's exploration is different from the trace of a human's design decisions.

Walk through each autonomy level. Notice the progressive shift: the engineer moves from doing the work to directing the work to overseeing the work. At each level, the required trust increases — and the mechanisms for establishing that trust must become more sophisticated. An AI-led agent needs to explain its exploration strategy. A fully agentic system needs to explain its reasoning about constraint modifications and strategy adaptations.

Assessment