Can subjective design judgment be turned into consistent grading criteria?

Rajasekaran describes a three-agent harness — planner, generator, evaluator — modeled on Generative Adversarial Networks, applied across two different domains: "one defined by subjective taste, the other by verifiable correctness and usability." The planner "took a simple 1-4 sentence prompt and expanded it into a full product spec," automating a step that "our previous long-running harness required the user to provide" upfront. The core move for the subjective domain was converting an ungradable question into a gradable one: "'Is this design beautiful?' is hard to answer consistently, but 'does this follow our principles for good design?' gives Claude something concrete to grade against."

The evaluator did not grade a static artifact. Given the Playwright MCP (a tool that lets the agent control a browser), "the evaluator would navigate the page on its own, screenshotting and carefully studying the implementation before producing its assessment," then wrote a critique that fed back to the generator for the next attempt. Rajasekaran ran "5 to 15 iterations per generation," with full runs stretching "up to four hours" because each cycle involved live browser interaction rather than scoring a screenshot, and instructed the generator to either refine a trending direction or "pivot to an entirely different aesthetic if the approach wasn't working." Calibration mattered: the evaluator was tuned with "few-shot examples with detailed score breakdowns" to align its judgment and "reduced score drift across iterations," and criteria explicitly "penalized highly generic 'AI slop' patterns." The same structure, Rajasekaran writes, "maps naturally onto the software development lifecycle, where code review and QA serve the same structural role as the design evaluator," carrying over two lessons from earlier harness work: "decomposing the build into tractable chunks, and using structured artifacts to hand off context between sessions."

The structured-artifact handoff echoes Can orchestration layers make coding agents more auditable?, where persistent state objects link decisions to execution across a session; here the same kind of artifact hands context between a planner, generator, and evaluator rather than between an agent and a human reviewer. It also bears on Can language models build and maintain their own agent harnesses?: where that paper argues a harness should be evaluated as infrastructure separate from the model, this source is a hand-built instance of that infrastructure, assembled and tuned by an engineer rather than produced by the model itself. Does self-editing through reviewed commits improve agent performance? sits in contrast: that harness evolves its own code under review, while this one keeps its three-agent structure fixed and iterates on the generated artifact instead. The leniency problem this architecture is built to route around is detailed in Can separating generator and evaluator agents reduce self-praise bias?.

The excerpt reports no benchmark scores, no comparison against the "ceilings" the earlier harness hit, and the "Planner" section is cut off before describing how planning integrated with the generator-evaluator loop in the coding domain. It is a first-person engineering account published on Anthropic's own blog, not an evaluated study, so the claim that this architecture "pushed Claude further" rests on the author's own framing rather than a reported metric. What the excerpt does establish is a specific, reusable harness design — subjective criteria made gradable, context handed off through structured artifacts, iteration bounded by wall-clock evaluator cost — whose generality beyond this one engineer's two projects is not shown.