News

What each tool shrinks, how you recover omitted content, telemetry defaults, and what the published numbers measure.

21 adapters for the agent you already built

Eight TypeScript adapters, thirteen Python adapters, and recoverable tool-result compression inside the agent you already run.

Seven levers, the measured number behind each, what breaks them, and the cases where they lose money.

Per-call api_base, fleet-wide config.yaml, Claude Code through both, and a routing callback that never touches inference.

Same six tasks, 54 runs, an exact-answer check on every one. 33.2 percent against 6.7, and the rows where each lost.

Four counters, one of them ten times cheaper than the others, and a prefix that decides which one you pay.

224,655 characters of tool schemas on one first turn. How to measure yours and three ways to bring it down.

65 percent fewer output tokens on ten prompts, 1,000 to 1,500 input tokens per turn to get it, and plans where it saves nothing.

The same two levers as Claude Code, the exact install for each agent, and the meters where token savings do not help.

Paired tasks, provider counters, a correctness gate, the spread, and four labels that stop an estimate becoming a claim.

A token counter said 96.2 million saved. Paired trials found a higher bill. The denominator explains both.

One real agent loaded 224,655 characters of tool schemas before its first action. Most came from plugins it might never call.

Compaction can preserve the job while deleting its rules. One benchmark found a 47-token buffer prevented the failure.

A 998-task replay gained seven net passes while 125 outcomes flipped. Aggregate scores hid both directions.

Nine quality-held pairs produced an attractive diagnostic and three publication blockers. The number stayed out of our claims.

The skill now runs at the API level of a green AI platform. No install, nothing to maintain. First platform integration.

The price of a token collapsed and the bill went up. What follows from reading that correctly.