Beyond Generating Code
AI has become increasingly capable of generating code, writing tests, and assisting developers with implementation. But a more significant opportunity is emerging: using AI agents not just to build a predefined solution, but to explore which solution should be built in the first place.
In complex engineering, the challenge is rarely limited to producing an implementation. Engineers must navigate competing constraints, evaluate alternative architectures, verify correctness, and balance performance against cost. Exploring these possibilities manually can require substantial time and expertise.
AI agent factories offer a different approach. By coordinating multiple specialized agents in an iterative workflow, engineering teams can generate alternatives, implement candidates, evaluate their performance, and use the results to guide subsequent experiments.
The potential shift is from AI-assisted implementation to AI-driven design discovery.
How Agent Factories Explore Engineering Designs
An agent factory is a coordinated system in which multiple AI agents contribute to a shared objective. Rather than relying on one agent to complete an entire task, the system distributes work across specialized agents and uses evaluation results to guide the next steps.
A typical engineering workflow involves five stages:
1. Define the design space. Engineers establish the objective, constraints, evaluation criteria, and acceptable trade-offs. These might include performance, resource consumption, reliability, or implementation complexity.
2. Generate design alternatives. Agents propose different architectures, algorithms, or implementation strategies. Each candidate represents a potential way to satisfy the defined requirements.
3. Implement candidates. Selected alternatives are translated into executable code, configurations, or other testable engineering artifacts.
4. Evaluate and verify. Automated tools assess candidates against measurable criteria. Depending on the engineering domain, this may involve simulation, testing, synthesis, or formal verification.
5. Guide further exploration. Results help determine which hypotheses deserve additional investigation. The system can prioritize promising directions, reject weak candidates, and refine subsequent proposals.
The important distinction is that evaluation becomes part of the generation process. Instead of producing a single answer and stopping, the system uses evidence from each iteration to inform what it explores next.
What Chip Design Reveals About This Approach
Chip design illustrates why iterative design discovery could matter. Hardware engineers must balance computational performance, resource constraints, correctness, and implementation feasibility. Small architectural changes can produce significantly different results, making systematic exploration valuable.
In one reported experiment involving hardware optimization for quantum error correction, a multi-agent workflow proposed architectural changes, implemented them, checked correctness, and used high-level synthesis results to guide further exploration.
Under a budget of 100 design hypotheses, the experiment reported a 52× speedup over the baseline in high-level synthesis estimates, with approximately US$26 in inference costs for the configuration using open models. A proprietary-model configuration reported a 204× speedup in the same estimation context, at approximately US$299 in inference costs.
These results are specific to the reported workload and experimental setup. High-level synthesis estimates should not be interpreted as equivalent to measured performance of a completed physical chip, and the figures do not establish that one configuration will outperform another across all engineering tasks.
The experiment nevertheless demonstrates an important principle: an agent-based workflow can use implementation and evaluation feedback to explore design alternatives, while the choice of models and inference infrastructure can materially affect the economics of that exploration.
It also highlights a practical limitation. Increasing the search budget does not necessarily produce proportional gains. In the same experiment, doubling the hypothesis budget from 100 to 200 increased the reported speedup from 52× to 60×. More exploration can help, but its marginal value must be measured.
Where Could Design Discovery Go Next?
The underlying approach could be relevant beyond chip design, although each application would require domain-specific validation.
- Mechanical engineering: Explore component geometries or structural configurations against strength, weight, and material constraints.
- Industrial optimization: Evaluate production schedules, resource allocation strategies, or process configurations against throughput and operating costs.
- Infrastructure engineering: Compare system architectures or network configurations against capacity, resilience, and efficiency requirements.
- Software architecture: Explore alternative service designs, data flows, and system configurations against latency, scalability, reliability, and cost targets.
These are potential applications, not proof that agent factories can autonomously solve every engineering problem. Their suitability depends on whether the design space can be represented adequately, candidates can be evaluated reliably, and the consequences of incorrect results can be controlled.
What It Takes to Make Agent-Driven Engineering Useful
Coordinating multiple agents is only one part of the challenge. Engineering teams need a disciplined framework to ensure that exploration produces useful, trustworthy results.
Reliable evaluation: Candidates must be judged against measurable criteria rather than subjective model preferences. Where possible, use independent testing, simulation, or formal verification.
Explicit constraints: Agents need clear boundaries defining acceptable solutions, including safety, compliance, resource, and operational requirements.
Controlled experimentation: Track hypotheses, implementations, evaluation results, and changes so that promising outcomes can be reproduced and investigated.
Efficient orchestration: Parallel agents can expand the search, but coordination overhead, inference costs, duplicated work, and diminishing returns must be considered.
Human oversight: Engineers should retain responsibility for defining objectives, reviewing high-impact decisions, and approving designs before deployment or physical implementation.
For CTOs, the business case should focus on whether the system can improve the quality of engineering decisions not simply how many agents it runs or how many alternatives it generates.
A practical assessment starts with five questions:
- Is the engineering objective measurable?
- Can candidate designs be evaluated automatically and reliably?
- Can the search space and unacceptable outcomes be clearly constrained?
- Are results reproducible and independently verifiable?
- Does the improvement justify the cost of experimentation and integration?
If these conditions are absent, adding agents may increase complexity without improving outcomes.
From Faster Implementation to Better Decisions
AI agent factories point toward a broader role for AI in engineering. Rather than serving only as assistants that generate code from human-defined specifications, coordinated agents could help teams systematically investigate design alternatives and learn from measurable evaluation results.
The opportunity is not unlimited autonomy. It is the ability to explore more possibilities within well-defined constraints, supported by reliable verification and human engineering judgment.
The next competitive advantage may not come from generating designs faster, but from discovering which designs are worth building.