The engineering design research community has studied agentic AI systems that use Large Language Model (LLM) agents to automate the engineering design process. However, these systems are prone to some of the same pathologies that plague humans. Just as human designers, LLM design agents can fixate on existing paradigms and fail to explore alternatives when solving design challenges, potentially leading to suboptimal solutions. In this work, we propose (1) a novel Self-Regulation Loop (SRL), in which the Design Agent self-regulates and explicitly monitors its own metacognition, and (2) a novel Co-Regulation Design Agentic Loop (CRDAL), in which a Metacognitive Co-Regulation Agent assists the Design Agent in metacognition to mitigate design fixation, thereby improving system performance for engineering design tasks. In the battery pack design problem examined here, we found that the novel SRL and CRDAL systems generate designs with better performance, without significantly increasing the computational cost, compared to a plain Ralph Wiggum Loop (RWL) Further, the novel CRDAL generates designs with significantly better performance than SRL. Also, we found that the CRDAL system navigated through the latent design space more effectively than both SRL and RWL. The proposed system architectures and findings of this work provide practical implications for future development of agentic AI systems for engineering design.
Figures & tables
Figure 1: A Common 18650 Lithium Ion Battery. 2 2 2 "Polymer Lithium Ion Battery - 18650 Cell (2600mAh, Solder Tab)" by SparkFunElectronics is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/?ref=openverse.
Figure 2: Example of a simple 6(W)×4(D)×2(H) battery pack composed of 18650 Lithium-ion battery cells using hexagonal close-packing. 3 3 footnotemark: 3
Figure 3: System flowchart of the Ralph Wiggum Loop (RWL).
Figure 4: System flowchart of the Self-Regulation Loop (SRL).
Figure 5: System flowchart of the Co-Regulation Design Agentic Loop (CRDAL).
Figure 6: Design capacity of the battery pack created by each agentic design system. p-value annotation legend: ns: 1.70×10−2<p ; ∗:1.00×10−2<p<=1.70×10−2 ; ∗∗:1.00×10−3<p<=1.00×10−2 ; ∗∗∗:1.00×10−4<p<=1.00×10−3 ; ∗∗∗∗:p<=1.00×10−4.
System
Success Rate
Range ( Ah )
Mean ( Ah )
Std. ( Ah )
Median ( Ah )
RWL
29/30
[50.00,135.00]
72.07
20.52
62.50
SRL
30/30
[50.00,192.50]
106.08
45.60
100.00
CRDAL
30/30
[50.00,187.50]
141.17
48.45
160.00
Table 1: Summary statistics on the design capacity of battery pack designs created by agentic design systems.
Comparison Group
T-test results
RWL vs SRL
t(40.6)=3.715,p=0.001, Cohen’s d=0.956
RWL vs CRDAL
t(39.4)=7.174,p<0.001, Cohen’s d=1.846
SRL vs CRDAL
t(57.8)=2.888,p=0.005, Cohen’s d=0.746
Table 2: T-test results on the design capacity in pairwise comparisons between agentic design systems.
Figure 7: Number of design steps taken before final design by each agentic design system. p-value annotation legend: ns: 1.70×10−2<p ; ∗:1.00×10−2<p<=1.70×10−2 ; ∗∗:1.00×10−3<p<=1.00×10−2 ; ∗∗∗:1.00×10−4<p<=1.00×10−3 ; ∗∗∗∗:p<=1.00×10−4.
System
Range (steps)
Mean (steps)
Std. (steps)
Median (steps)
RWL
[2,30]
9.07
7.02
6.50
SRL
[4,27]
9.10
4.45
8.00
CRDAL
[3,17]
9.37
3.20
9.00
Table 3: Summary statistics on the number of steps taken before final design by agentic design systems.
Comparison Group
T-test results
RWL vs SRL
t(49.1)=0.022,p=0.983, Cohen’s d=0.006
RWL vs CRDAL
t(40.6)=0.213,p=0.832, Cohen’s d=0.055
SRL vs CRDAL
t(52.7)=0.266,p=0.791, Cohen’s d=0.069
Table 4: T-test results on the number of steps taken before final design in pairwise comparisons between agentic design systems.
Figure 8: Design step trajectory and final design in the latent design space, explored by each agentic design system.
Figure 9: Cell series and parallel connections explored by each agentic design system.
Dresden University of Technology, Dresden, Germany · MAN Truck & Bus SE, Munich, Germany · German Research Center for Artificial Intelligence, Kaiserslautern, Germany +1