This post reports work on two problems the author treats as interconnected: getting Claude to produce high-quality frontend designs, and getting it to build complete applications without human intervention. Earlier efforts on a frontend design skill and a long-running coding agent harness had improved performance above baseline through prompt engineering and harness design, but the author reports both hit ceilings — which is given as the reason for seeking an approach that would hold across two domains as different as subjective taste and verifiable correctness.
The approach taken is borrowed from Generative Adversarial Networks: a multi-agent structure pairing a generator with an evaluator. The post’s central claim about why this helps is about self-evaluation rather than about generation — agents asked to judge their own work tend to praise it confidently even when quality is visibly mediocre, and the author argues that tuning a standalone evaluator to be skeptical is far more tractable than making a generator critical of its own output. Separation does not by itself remove the leniency, since the evaluator is still an LLM inclined to be generous toward LLM output, but it gives the generator something concrete to iterate against.
Applied to long-running coding, this became a three-agent planner–generator–evaluator architecture producing full-stack applications over multi-hour autonomous sessions. The later half of the post is an account of taking that architecture apart again as models improved, and the closing position is that the space of interesting harness combinations does not shrink as models get better — it moves.
Key Points
- Two failure modes are identified in agents on long tasks: loss of coherence as the context window fills, and Context Anxiety, where a model wraps up work prematurely as it nears what it believes is its context limit. Context Reset is presented as addressing both.
- A reset is distinguished from Compaction on the grounds that compaction summarises history in place and so preserves continuity without giving the agent a clean slate, which means context anxiety can persist through it; the cost of a reset is that the handoff artifact must carry enough state, plus added orchestration complexity, token overhead and latency.
- For frontend design the author wrote four grading criteria given to both generator and evaluator: design quality (whether the design reads as a coherent whole), originality (evidence of custom decisions rather than template layouts and library defaults), craft (typography hierarchy, spacing, colour harmony, contrast ratios) and functionality (usability independent of aesthetics).
- Design quality and originality were weighted above craft and functionality, on the stated grounds that Claude already scored well on the latter two by default while producing bland output on the former two; the criteria explicitly penalised generic patterns, naming purple gradients over white cards as a telltale sign of AI generation.
- The evaluator was calibrated with few-shot examples carrying detailed score breakdowns, which the author reports aligned its judgment with his preferences and reduced score drift across iterations.
- The frontend loop was built on the Claude Agent SDK, with the evaluator given the Playwright MCP so it could navigate and screenshot the live page before scoring rather than grading a static image; runs of 5 to 15 iterations stretched up to four hours of wall-clock time.
- The wording of the criteria shaped output in ways the author reports not fully anticipating — including the phrase “the best designs are museum quality”, which pushed designs toward a particular visual convergence.
- Score improvement across iterations is reported as real but not cleanly linear: the author regularly preferred a middle iteration to the last, and notes that even the first iteration beat an unprompted baseline, suggesting the criteria steered the model before any evaluator feedback arrived.
- One reported run illustrates a discontinuous change: asked for a Dutch art museum site, the model produced a polished dark-themed landing page by the ninth iteration, then on the tenth scrapped it entirely for a 3D room with a CSS-perspective checkered floor, artwork hung in free-form positions and doorway-based navigation instead of scrolling or clicking.
- In the full-stack harness the planner expanded a 1–4 sentence prompt into a full product spec; it was instructed to be ambitious about scope and to stay at the level of product context and high-level technical design, on the stated reasoning that granular technical details specified upfront would cascade errors into implementation.
- The generator worked in sprints, one feature at a time, on a React, Vite, FastAPI and SQLite (later PostgreSQL) stack with git for version control; before each sprint the generator and evaluator negotiated a sprint contract agreeing what “done” meant and how it would be verified, with the two communicating by writing and reading files.
- On a retro game maker prompt, a solo run took 20 minutes and 200 — over 20x the cost. The solo run’s application had broken wiring between entity definitions and the game runtime, so the game did not respond to input; the harness run’s spec ran to 16 features across ten sprints and produced a playable result.
- The author reports Claude is a poor QA agent out of the box, watching it identify legitimate issues and then talk itself into approving the work anyway, and testing superficially rather than probing edge cases; the fix described is reading evaluator logs, finding where its judgment diverged from the author’s, and updating the QA prompt over several rounds.
- A general principle is drawn for harness design: every component encodes an assumption about what the model cannot do on its own, and those assumptions are worth stress testing because they may be wrong and because they go stale as models improve. The post quotes Anthropic’s own Building Effective Agents post for the underlying idea — “find the simplest solution possible, and only increase complexity when needed”.
- Context resets were not part of this harness to begin with: Opus 4.5 had largely removed context anxiety on its own, so the agents were run as one continuous session with the Agent SDK’s automatic compaction handling context growth.
- The later simplification pass proceeded by removing one component at a time and reviewing the impact of each, after a first attempt at cutting the harness back radically failed to replicate the original’s performance and left it hard to tell which pieces were load-bearing. The sprint construct was removed after Opus 4.6 landed, and the evaluator moved to a single pass at the end of the run rather than grading per sprint.
- The evaluator’s value is reported as conditional rather than fixed: as raw model capability rose, tasks that previously needed its check came within what the generator handled alone, so the post’s stated rule is that an evaluator is worth its cost when the task sits beyond what the current model does reliably solo.
- The updated harness built a browser DAW in about 3 hours 50 minutes for 0.46, three build rounds of 2h07, 1h02 and 10.9 minutes, and three QA rounds of 8.8, 6.8 and 9.6 minutes.
- QA still found real gaps in that run, quoted in the post: first-round feedback that core DAW features were display-only without interactive depth, and second-round feedback that audio recording remained stub-only and clip resize and split were not implemented.
Context
This is a firsthand engineering account from one member of Anthropic’s Labs team, describing his own experiments rather than a controlled study. There is no baseline beyond the single solo-versus-harness comparison on one prompt, no repetition, and the quality judgments driving the frontend work are explicitly the author’s own taste — a point the post makes itself in noting that the evaluator was calibrated to align with his preferences.
The post is candid about limits in its own results. The harness-built game maker retained workflow problems the author reads as a gap in the base model’s product intuition rather than something the harness addressed; the generated game’s physics had rough edges, with his character jumping onto a platform and ending up overlapping it; and an AI-generated level left him stuck behind a wall. For the DAW he notes the application is far from a professional music production program and that Claude cannot hear, which made the QA feedback loop less effective on musical taste. Even the tuned evaluator is described as leaving small layout issues, unintuitive interactions and bugs in deeply nested features undiscovered.
Its relationship to the author’s earlier harness work is one of revision rather than extension: several components that earlier work established as necessary are removed here, and the stated reason is that the model changed underneath them.