Anthropic broke an agent ceiling by making "is this design good?" a gradable question

The order of operations is the part that transfers. The evaluator agent could not exist until someone had written down what a good design is in terms a grader can actually apply, which is why the post describes developing the criteria first and the agent second ^[5]. Architecture routes judgment; the judgment itself still has to come from written criteria.

The reason a rubric is load-bearing sits in the failure it corrects. Agents asked to grade their own output tend to praise it confidently even when a human reader would call the work mediocre, and the effect is worst on subjective tasks like design, where there is no binary check equivalent to a passing test ^[17][18]. Even on tasks with verifiable outcomes, the post reports agents showing poor judgment mid-task in ways that hold back performance ^[19]. Separating the worker from the judge is described as a strong lever, and in the same breath as a partial one, because the evaluator is itself a language model and does not become strict by being relabelled ^[20].

The cost of coherence is stated plainly. Clearing the context window and starting a fresh agent with a structured handoff fixes both the drift and the premature wrap-up, and it adds complexity to orchestration, spends extra tokens, and slows every harness run ^[13][16]. There is also role inflation to price: the earlier experiment ran two agents, an initializer that turned a product spec into a task list and a coding agent that implemented features one at a time, while the published design runs three ^[8][7]. That is one more role per build, half again as many as before, and each of them bills inference ^[21].

One reading of this account is that a model vendor's engineer is reporting that more of the vendor's inference fixes the vendor's plateau. What the post volunteers cuts against that reading. It says compaction was insufficient specifically because Claude Sonnet 4.5 exhibited context anxiety strongly enough to require full resets, which the post treats as a specific weakness in a specific model, distinct from a claim of superior capability ^[15]. It also concedes that resets buy their fix at a cost ^[16]. What it does not do is put a number on the improvement: the result is that the three-agent harness produced rich full-stack applications over multi-hour sessions, with no benchmark or measured baseline comparison in the account ^[22].

So the quarter's decision is smaller and harder than a diagram. If your agents plateau, the work is not adding a grader; it is writing the criteria the grader will use, and then owning them. Decomposition into tractable chunks and structured handoff artifacts were already carried forward from the earlier harness ^[6], and the rubric now joins them as a document the pipeline depends on. Next quarter's failure mode follows from that: an evaluator inherits whatever definition of quality you gave it, so a stale or wrong rubric produces confident approval at scale, which is the same problem the separation was meant to solve, moved one layer up.