A coding agent's bill tracks its cache, not the tokens you trim
Each of tonight's four developments lands one layer below its launch tweet, in the mechanism a builder operates rather than the announcement they saw.
Cutting a coding agent's tokens barely moves its bill1
The paper reconstructs cost on 2,848 Claude Code runs across 103 tasks and splits the bill into four components. Prompt caching, not tool-output tokens, accounts for about 87% of that reconstructed cost. Per-task token reduction barely tracks billed cost (Pearson r=0.15), and one compression method removed an estimated 38% of tokens while raising cost 6.8% and dropping patch success from 27 of 40 tasks to 15 of 40.1
Why it matters: the meter a builder tunes against is cache traffic, so pruning tool output to save money can raise the bill and corrupt the context the agent needs to finish the task. This rests on a single July 13 preprint with no independent replication yet, so read it as one well-instrumented result rather than a settled law. It still targets the exact lever most teams pull to cut agent costs, and it measures that lever pointing the wrong way.
Dependabot now holds a new release three days before opening a version-update PR2
Dependabot waits until a package has sat on its registry for at least three days before it opens a version-update PR, now the default across every supported ecosystem on github.com and requiring no configuration; security updates still open immediately. GitHub's reasoning is that a compromised or broken release often reaches dependency updates before maintainers and the community have caught it.2 The window governs only when Dependabot proposes a change, not what a package manager will install: a manual update or a regenerated lockfile can still resolve a package younger than three days, and default release-age gates run shorter (npm none, pnpm 10 zero, pnpm 11 and Yarn one day).3 Writing in January, before the default flip, security researcher Christian Schneider documented 2026 supply-chain incidents caught and removed within 24 to 72 hours and argued a seven-day cooldown clears that window every time.4
Why it matters: three days sits at the low edge of the detection window Schneider measured, and it covers only the surface Dependabot controls. Set the same age gate in your package manager and CI, or the new default guards less of the pipeline than the headline suggests.
New York enacts the first statewide hyperscale data-center moratorium5
Governor Kathy Hochul signed an executive order pausing discretionary state permits not already deemed complete for new hyperscale data centers, for up to a year while the state develops a Generic Environmental Impact Statement. The order directs the Department of Public Service to make data centers either pay more for their energy or supply their own, and gives Empire State Development 60 days to issue a Community Investment Framework.5 Reporting pins the threshold at 50 megawatts or more, a figure the governor's release does not state, and notes the order differs from a one-year moratorium the legislature passed last month that Hochul has not signed.6
Why it matters: for anyone siting US inference or training capacity, New York is now the first state where a hyperscale build can stall a year on a discretionary permit, and the price of grid power is being written into that permit. This is an executive pause pending a framework, not the legislature's separate bill, so its terms can shift as the impact statement takes shape.
Cloudflare's Precursor scores the whole session, defeating the refresh reset7
Precursor injects a compact, per-response JavaScript bundle that records pointer movement, keyboard cadence, focus changes, and visibility events, then accumulates the signal across a whole session instead of scoring each request on its own. Session scoping means a bot cannot reset its behavioral signature by refreshing the page. The tell is motor behavior over time, where humans produce wrist-pivot arcs, cognitive delays, and hand-tremor oscillation while automation moves in linear interpolations or mathematically ideal Bézier curves that diverge from a human's the longer a session runs.7 Cloudflare says it analyzes over a trillion requests a day across more than 20% of the web, that automated traffic is now roughly 57% of web requests, and that Precursor is free until its GA release later this year.7 An independent write-up flags the tradeoff the launch asserts away: this behavioral telemetry, once confined to login and challenge pages, now runs across ordinary browsing, with disclosure left to each site operator.8
Why it matters: if you run a browsing or agentic client, the evasions that beat per-request challenges (solve once, refresh, mimic a human for a burst) stop working against a signal that scores the session. Deploy it on your own site and you have widened your client-side tracking surface, which becomes a privacy-notice update rather than a checkbox.
Sources
- arXiv · Weinberger & Hozez, "Token Reduction Is Not Cost Reduction"
- GitHub Changelog · Dependabot version updates introduce default package cooldown
- TECHi · GitHub's Dependabot cooldown has an install-layer gap
- Christian Schneider · Dependency cooldowns: a simple supply chain fix
- Office of Governor Kathy Hochul · First statewide moratorium on new hyperscale data centers
- NY1 · Hochul enacts first statewide data center moratorium
- Cloudflare · Introducing Precursor: detecting agentic behavior with continuous client-side signals
- CyberInsider · Cloudflare expands behavioral tracking to fight AI bots