You trimmed the logs; the cost was in the turns

Three papers about the plumbing around the model call: the agent loop whose bill is set by how many times it goes round rather than how much it reads back, the second opinion that costs nearly three times as much and rarely beats the better model on its own, and the parameter schema that quietly decides which of your tools gets called.

Today's pick

84% vs 4%

share of the variation in a coding agent's trajectory cost explained by its turn count, against the average size of its tool output

TurnCost: Turns, Not Observations: Cache-Aware Loop Engineering for Coding Agents

Trimming what your tools print back is the obvious way to cut an agent's bill, and on 600 public coding-agent trajectories it is close to the wrong lever: the context is append-only, prefix caching served 98.6% of the input tokens, and so pricing those tokens at the list input rate overstates the real cost about sevenfold. Fitting cost against each candidate driver, turn count explained 84% of the variation and mean observation size 4% — the most expensive runs were simply the longest, not the ones reading the biggest logs. The paper also derives when handing a large tool output to a cheaper model pays for itself, and the horizon is set by what cached reads cost rather than the list price, which is why delegation raised cost in about half the settings tested and plain truncation matched or beat it in all twenty threshold-and-setting combinations. Count turns before you compress logs, and price any cheap-model routing against your cached-read rate.

open the explainer → abstract PDF

9 of 96

review pipelines, out of every pairing tested, that beat simply calling the stronger of the two models by itself

Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines

Adding a second model to review the first one's answer looks like a clear win when you measure it against the first model — on one pairing it added 33.2 points — but most of that gain is the reviewer supplying answers it could have produced alone. Across 48 model pairs run in both directions on maths, knowledge, code and reasoning, review beat simply calling the stronger of the two by itself in only 9 of 96 directions, finishing about 3.1 points below it on average while costing 2.6 to 2.8 times as much. The mechanism is deference: where the first model was right and the reviewer wrong, the final answer stayed right 76% of the time, but where only the reviewer was right it kept its own answer just 40% of the time. A cheap selector over the same two calls — both models answer, a small classifier picks the likelier one — beat review by 2.7 to 4.2 points for less money, so measure your stronger model alone before you ship a critique pass.

open the explainer → abstract PDF code

19 of 30

model-and-task conditions where the parameter schema moved tool choice more than the server name, tool name or description did

Which Metadata Matters for LLM Tool Selection in MCP? Evidence from Controlled Factor-Level Interventions

Two tools can do the same job, carry the same name and still be picked at very different rates, and the deciding factor is the shape of their parameters. Across 367,174 published tool definitions the authors found 169 records named send_email exposing 78 distinct parameter sets, then ran controlled swaps on five models and five tasks, changing one piece of metadata at a time with candidate order rotated: the schema was the strongest effect in 19 of 30 model-and-task conditions, about twice the size of the server-name effect, and models consistently preferred separate top-level fields over one grouped object. Descriptions that spell out what each parameter does helped as well; tool names barely mattered, and where candidates differed only in their names, list position decided it 79% to 96% of the time. Renaming a field mattered as much as adding one: a tool whose recipient field was called recipient_id was chosen in 1% to 10% of trials, against 35% to 51% when the same field was called receiver.

open the explainer → abstract PDF