Yesterday's digest was about the bill for building on someone else's infrastructure. Today's is about the bill for delegating to agents — and it comes due in three different currencies: money, review time, and blast radius. One of the four stories is genuinely good news, so we'll start there.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Anthropic cancelled the Sonnet 5 price increase — check whether your September budget knows
Claude Sonnet 5 launched at $2 per million input tokens and $10 per million output, labelled introductory pricing that would expire on August 31 and step up to $3/$15 on September 1. That step-up is off. Anthropic's own pricing docs now carry a note stating plainly that the introductory rate "is now the standard price" and that the scheduled increase "will not occur." A 50% jump that a lot of people penciled into next month's forecast has simply evaporated.
Worth pausing on how quietly this landed. Several pricing trackers and cost calculators still list the August 31 deadline as live, which is a small lesson about where to get your numbers: the vendor's own docs page beat the aggregators by a wide margin here. Put this next to yesterday's DeepSeek story and you get one week in which two providers moved their prices in opposite directions, neither with much warning.
For builders
Open whatever spreadsheet or Notion doc holds your September inference forecast and delete the Sonnet line increase — that's real money you can stop reserving. Then do the reverse check on everything else you call: grep your provider docs, not a blog roundup, for the word "introductory." Every model you adopted this year at a launch discount has a date attached to it somewhere, and you'd rather find it now than in an invoice. Batch API on Sonnet 5 is $1/$5 per million, so anything of yours that isn't interactive is still leaving half on the table.
Story II
AI now writes just under half of everything in Linear — and none of it can get reviewed
Linear published its own usage data this month, and the headline number is stark: teams now use AI to write just under half of everything created in Linear. Two years ago that share was under 0.1%. Pull requests per workspace are up 111% between June 2024 and June 2026, and teams running coding agents tripled their weekly PRs from 21 to 65 while teams without them crept from 8 to 10. Product managers attaching pull requests went from 3% to 10%; designers from 1% to 8%. More people are shipping code than ever.
Then there's what happens to that code. LinearB's 2026 benchmarks, drawn from 8.1 million pull requests across 4,800 engineering teams in 42 countries, found that AI-assisted PRs wait about 5.25 times longer than unassisted ones just to get picked up for review — roughly 1,000 minutes versus 200. They merge 32.7% of the time against about 84.5% for unassisted code. And they're bigger: 400+ lines at the 75th percentile versus 157. The bottleneck isn't the reviewing; once a human actually starts, AI-assisted PRs get reviewed faster (194 minutes versus 252). The bottleneck is the queue in front of it. Linear's own report notes the same paradox from the other side — total time spent on product development went up, not down.
The uncomfortable reading is that a lot of this output is not slower work. It's work nobody wanted. A PR that sits ten times longer and then merges a third of the time was, statistically, a waste of somebody's afternoon.
For builders
Cap the size of what your agent hands you. That 400-line P75 is the single most fixable number in this story — instruct the agent to open one PR per logical change and stop when a diff crosses roughly 200 lines, and you move your own PRs out of the category that reviewers avoid. If you work solo, the equivalent discipline is refusing to merge anything you haven't read end to end; the 32.7% merge rate is a market signal that a meaningful share of agent output shouldn't ship. And if you're measuring yourself on PRs opened, switch to PRs merged this week. The first number is now trivially easy to inflate and means almost nothing.
Story III
Eight open-source agents ran a four-day intrusion on Taiwan's government
Between July 1 and July 4, suspected China-linked operators ran what researchers describe as a near-autonomous intrusion against Taiwanese government systems — 12 attack waves, up to eight sub-agents working in parallel, each assigned its own objective and technique. The haul: 85 compromised government accounts, more than 2,564 personnel records, seven SSO client secrets, and six internal database credentials spanning MSSQL, Oracle, and Sybase. Taiwan's nuclear safety agency, government email systems, supply-chain vendors, and seven energy companies were all reached. The Israeli firm Dream surfaced it from a 160 MB archive of 1,395 files documenting the operation, and the Financial Times reported it on August 12.
Here is the part that belongs in a builders' digest rather than a security newsletter. The framework was assembled from Hermes, an open-source agent framework, and OpenClaw, the open-source personal AI assistant that has been near the top of everyone's GitHub feed since January. No proprietary model, no enterprise account, no clearance. The agents researched, self-corrected, and switched tactics when a route failed. Humans built and pointed the thing; the four days of work in between were largely not theirs.
OpenClaw's security record makes this less surprising than it should be. An academic analysis of its advisory corpus counted 470 security advisories filed against the project between January 31 and April 15 alone, with the gateway WebSocket interface accounting for 121 of them — 7 critical, 47 high. That is what a framework looks like when adoption outruns its disclosure process.
For builders
If you run an agent framework locally — OpenClaw, Hermes, or any of the dozen wrappers around them — treat it as production infrastructure, because it holds shell access, browser control, and your credentials in one process. Three things worth doing this afternoon: confirm nothing is listening on an unauthenticated local port (the gateway interface is where the advisories cluster), give the agent a scoped API key with its own spend limit rather than your main one, and check what skills or plugins it has permission to install without asking you. The Taiwan operators didn't exploit anything exotic. They ran the software as designed, at a target that hadn't assumed anyone would.
Story IV
Grok 4.6 is the cheap seat at the frontier — until your agent's context crosses 200K
SpaceXAI released Grok 4.6 on August 12 with a 500,000-token context window and explicit tuning for long-running agents. It scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max, at $2 per million input and $6 per million output — cheap for that tier. It's available through the xAI API as grok-4.6, in Cursor and Grok Build, and routable via OpenRouter, Vercel, and Cloudflare. There's no open-weights release and no self-hosting path.
The number most people will miss: above 200,000 prompt tokens, pricing doubles to $4/$12. For a chat app that threshold is theoretical. For the long-running agent this model was built for, it is not — context accumulates by design, so a session that starts in the cheap tier finishes in the expensive one, and your cost per task depends on how long the agent runs rather than how hard the task was.
For builders
Before you switch an agent workload to any long-context model, find the tier threshold in its pricing page and log your actual prompt-token count per turn against it. If your sessions routinely cross 200K, the honest comparison isn't $2/$6 — it's $4/$12, and that changes which model wins. The cheap fix is context hygiene: summarize and drop completed tool output instead of carrying the full transcript forward, and cache the stable prefix (Grok 4.6 prices cached input at $0.50 per million below the threshold). An agent that forgets what it no longer needs is not just cheaper — going by Story II, it also produces smaller diffs.
— The Vibe Gate news desk. We read the firehose so you can keep building.