Allocation Beats Abundance

If You Only Read One Thing

Capability is starting to look less like a possession than a budget, and budgets only matter when someone allocates them well. Cadences Makes The Surface Count shows Claude Code changes delegation without changing the model, while Model Routing Hits The Tail uses co-failure math to show why model committees stop paying off when their errors line up.

Cadences Makes The Surface Count

Anthropic's newest usage report is nominally about economic rhythms. The more useful technical finding is that the product surface changes how much work users delegate to the same model.

The Cadences report, published June 26, expands Anthropic's Economic Index with higher-rate sampling, output classifiers, and monthly splits across Claude conversations and first-party API traffic. It finds that 93% of Claude conversations produce an artifact: explanations are 17% of conversations, documents and reports 15%, guidance 11%, and code or technical work about a sixth. Work conversations are disproportionately document-shaped, while app-building and similar artifacts consume far more compute than simpler explanations.

The important shift is delegation. Anthropic measures AI autonomy on a 1-5 scale and finds Claude Code has higher autonomy than chat or Cowork for 26 of 31 output types. Across all conversations, the average gap is 0.37 points; for scripts and code snippets it is 0.53 points. The easy objection is that Claude Code uses stronger models more often: 54% of Claude Code sessions are served by Opus, versus 10% of chat and Cowork. Anthropic tries to control for that and still finds a 0.26-point autonomy gap among Sonnet conversations.

Why it matters: This is the cleanest evidence yet that "agent" is not a model property. It is a model plus a surface that changes the shape of user intent, tool authority, feedback, and stopping conditions. Anthropic's example is blunt: a median chat or Cowork conversation producing a blog post has 13 rounds of back-and-forth, while a median Claude Code session producing a blog post has a single human prompt. The same category of output becomes more delegated because the interface invites execution rather than discussion. That changes how to read model benchmarks. A model that looks weaker in chat can do more useful work in a harness that gives it files, tools, persistence, and a clear artifact boundary.

The cost finding points the same way. Higher-wage mapped tasks consume more tokens, and Anthropic says about 44% of that wage gradient is explained by output mix. More valuable tasks are not simply shorter because the model is smarter. They are longer because the human and model are negotiating a more complicated artifact.

Room for disagreement: This is Anthropic telemetry on Anthropic products, not a neutral market census. The user base is self-selected, and the artifact classifier is only as good as its labeling system. Still, the same-model comparison is hard to dismiss: product design appears to move autonomy even when model choice is held closer to constant.

Model Routing Hits The Tail

The default engineer answer to model uncertainty is to add another model. A new paper argues that this helps only when the models fail on different questions.

Josef Chen's co-failure paper, submitted June 25, studies routing, voting, cascades, fusion, and mixture-of-agents across 67 models from 21 providers. The key variable is beta, the all-wrong rate: how often every model in the pool fails on the same query. For any policy that selects one member model's answer, accuracy cannot exceed one minus beta. The familiar pairwise error-correlation number is not enough, because two model pools can look similarly correlated while having very different all-wrong tails.

The empirical results are the useful part. On open-ended math, observed beta is 0.052 versus 0.023 under the paper's full 67-model statistical baseline, about 2.5 times underpriced. On execution-graded code, beta is 0.079. Re-asking GPQA-Diamond, a graduate-level science Q&A benchmark, in free-response form rather than multiple choice pushes beta to 0.127, with grading from a five-judge LLM panel. At matched quality, heterogeneous low-correlation ensembles beat high-correlation self-mixtures, but the paper's blunt conclusion is that on checkable tasks, combining models rarely beats the single best model without a strong query-level routing signal.

Why it matters: The practical lesson is that orchestration is allocation, not magic. A router can harvest resolvable disagreement: one model is wrong, another is right, and the system can identify which answer to trust. It cannot harvest shared ignorance. That matters for agent frameworks because "use several models" is becoming the lazy substitute for understanding the task. If the failure mode is common across the frontier pool, adding more model calls mainly adds cost, latency, and a false sense of redundancy.

The strongest connection is to last week's vLLM Semantic Router Fusion story. Model panels inside a serving router sound like a free quality layer. Chen's result says the panel only has economic value if the system knows which questions create differentiated failures. Otherwise, the single best endpoint plus a good verifier may beat a committee that fails together.

Room for disagreement: The paper's own dataset card states the limits: GPQA free-response grading uses an LLM-judge panel, code grading is strict but unofficial, and generations are at temperature zero. Systems that generate new evidence through tools, tests, retrieval, or human review can move outside a pure answer-selection ceiling. That is the point, though: the escape hatch is better verification and task evidence, not more undifferentiated model opinions.

The Contrarian Take

Everyone says: Agents need stronger models and smarter orchestration. More choices in the model pool should mean better answers.

Here's why that's wrong (or at least incomplete): Today's evidence points to allocation, not abundance. Cadences shows that the same model behaves more autonomously when the product surface changes the work contract. The co-failure paper shows that model committees stop helping when every member fails on the same query. The durable advantage is not calling more models. It is putting judgment in the right place: interface, router, verifier, or human checkpoint.

Under the Radar

Cline rolled stable back after the SDK jump. Cline 4.0.1 ships the 3.89.2 extension code under a higher version number after regressions in 4.0.0, while SDK-migration work continues separately on main. That is a production signal: agent runtimes cannot migrate their internal SDK layer faster than their extension surface can absorb.
Vercel pushed service tiers into provider options. The latest Vercel AI SDK Fireworks package adds a serviceTier provider option for Fireworks chat models. It is small, but it shows latency/cost class becoming an SDK-level routing knob rather than an out-of-band provider setting.

Quick Takes

Aider is still a model-adapter shock absorber. Aider v0.85.0 added support for newer Gemini 2.5 models, Responses API models, thinking-token controls, and higher DeepSeek token limits. The point is not any one model ID; coding tools are becoming compatibility layers that hide provider churn from the editing loop. (Source)
llama.cpp moved OpenCL attention work forward. Release b9828 reworks OpenCL flash-attention kernels, adds prefill prepass kernels, pads KV and mask tiles, and can skip fully masked tiles. Yesterday's CUDA synchronization story was not an outlier: local inference gains are increasingly kernel scheduling and memory layout work. (Source)
Artificial Analysis is making cost decomposition harder to ignore. Its homepage now breaks task cost into answer, reasoning, cache write, cache hit, and non-cache input components, and distinguishes cache pricing by provider. That is the right accounting unit for agents, where repeated context can dominate the bill. (Source)

The Thread

Today's issue is about the limits of "just add capability." Cadences says the interface decides how much capability gets delegated. The co-failure paper says model orchestration has a mathematical ceiling unless failures are genuinely diverse and detectable. Cline, Vercel, Aider, llama.cpp, and Artificial Analysis are smaller versions of the same shift: production AI is becoming an allocation problem. The scarce resource is no longer only model intelligence. It is knowing where to spend judgment, tokens, latency, and trust.

Predictions

New predictions:

I predict: By August 31, 2026, at least one public coding-agent benchmark or lab usage report will add a delegation/autonomy metric that separates model capability from harness or product surface. (Confidence: medium; Check by: 2026-08-31)
I predict: By September 30, 2026, at least one public model-routing benchmark or router product will report an oracle upper bound, all-wrong rate, or equivalent co-failure metric alongside accuracy and cost. (Confidence: medium; Check by: 2026-09-30)

Generated: 2026-06-28 03:40 EDT