Three new cards, three lanes. Tunguz on the infrastructure economics of agent token demand. Willison on coding agents as practitioners of professional visual tools. Indig on what marketing teams are not counting when they claim AI saves time.
The parallelization economy
On February 6, 2026, agents consumed more tokens on OpenRouter than humans did. Tunguz published a framework on September 7 showing that date was not a curiosity. It was a structural crossing.
Agent token consumption crossed human consumption on February 6, 2026 and now follows three order-of-magnitude waves driven by parallelization, not generation speed names three waves. Wave one is chat, roughly one million tokens per active user per day. Wave two is single agents, 100 to 200 million tokens per agent daily. By June 2026, Codex alone accounted for 64% of combined enterprise token usage. Wave three is meta-harnesses dispatching parallel sub-agents, consuming billions of tokens per run. One example from the post: analyzing 200 YCombinator startups across 30 data columns costs tens of millions of tokens per execution.
"Consumption does not grow because generation gets quicker. It grows because parallelization compounds."
Goldman Sachs projects 120 quadrillion tokens per month by 2030, 24 times the 2026 level. The projection is speculative. The underlying mechanism is not: every layer of agent nesting multiplies token volume, and the multiplier compounds with each architectural step from chat to single agents to meta-harnesses.
This card pairs with Reported 3x AI productivity gains are agent shift-work economics, not intelligence multiplication from September 8, which interrogates whether the 3x productivity gain from that same consumption is real or accounting. Together they describe what the surge is and what it actually produces.
Agents as practitioners of professional tools
On September 5, Willison documented a working render-inspect-edit loop using Blender from natural language. Giving a coding agent access to a professional visual tool with a Python API enables autonomous render-inspect-edit iteration without human direction between rounds captures the mechanism: a frontier coding agent given the application path and a natural language target writes Python against Blender's scripting API, triggers a render, inspects the output image, and revises the Python for the next pass. The human provides the initial prompt. The agent drives every round after that.
"Modern frontier models have got really good at using Blender."
The pelican-on-a-bicycle scene Willison tested cost approximately $4.24 at equivalent gpt-6-astra API pricing. The practical reach of this pattern extends to any professional tool that exposes a Python API covering the full scene or document graph. Blender is the test case. The class of applicable tools is larger.
The limitation Willison does not resolve: a flat render per iteration may not surface geometry errors or camera problems that a turntable view would catch. Production use requires more stringent inspection than creative experimentation demands.
The hidden tab in AI productivity
Kevin Indig's September 7 post for Growth Memo names what goes uncounted when marketing teams report AI time savings.
AI tool maintenance in marketing teams creates invisible overhead that the official hours count does not capture, making net productivity gains structurally smaller than they appear synthesizes four independent studies. METR measured software developers expecting 24% faster performance and finding 19% slower. Workday measured 4 of every 10 AI-saved hours returning as correction and rewriting. BetterUp and Stanford found 41% of team members encountered workslop in the previous month at an average 1 hour 56 minutes per incident to resolve. HubSpot found 66% of companies are building internal AI tooling for their marketing teams.
The mechanism: teams count what the tool produces, not what maintaining and correcting the tool costs. Prompting, validating, rewriting, and updating workflows all consume the same hours that would otherwise go to brand authority, digital PR, and earned media. Those hours are slow to compound but they are the only path to trust signals that AI content cannot purchase.
"AI adoption can turn marketers into unpaid (ineffective) software teams, while the real work absorbs the opportunity cost."
The actionable frame Indig offers: the skill worth developing is the ability to name and defend when a specific AI workflow is saving time versus when it is draining time from the core brand-building work. Not all AI use in marketing is equal. The ones worth keeping are the ones that survive a rigorous accounting of the maintenance hours.
- Tom Tunguz, Agent token consumption crossed human consumption on February 6, 2026 and now follows three order-of-magnitude waves driven by parallelization, not generation speed. Agent token consumption crossed human consumption on February 6, 2026, and now follows three waves at 1M, 100-200M, and billions of tokens per day, each driven by parallelization.
- Simon Willison, Giving a coding agent access to a professional visual tool with a Python API enables autonomous render-inspect-edit iteration without human direction between rounds. A coding agent given a natural language prompt and access to Blender via its Python API can drive the full render-inspect-edit loop without human intervention between rounds.
- Kevin Indig, AI tool maintenance in marketing teams creates invisible overhead that the official hours count does not capture, making net productivity gains structurally smaller than they appear. AI tool maintenance in marketing creates invisible overhead that makes net productivity gains structurally smaller than official counts show; the hidden cost displaces compounding brand-authority work.