The Week in AI Adoption: Who Owns What the Agent Did, May 22
Across adoption, engineering, markets, and the self, this week's 28 posts kept circling one unanswered question: who owns what the agent did?
The week in one glance
- Boards are being asked to name an owner for every agent in production, every dollar of agent spend, and every regulatory bet before Q3 close.
- Vendors swapped models and cut prices under teams that never chose the change, so the harness question became an ownership question.
- AI agents now place real orders, and the edge, the audit, and the data all expire faster than the agent that uses them.
- Four technostress studies converged on one quiet cost: the blurring of which work, which ideas, and which skill are still distinctly yours.
Theme of the week
Twenty-eight posts went out this week across all four Cerevisor tracks, and they kept arriving at the same sentence with different nouns: when an agent acts in your name, who owns the result? On the adoption side, two agent-identity surveys found that no company can name every agent running in production. In engineering, GitHub swapped the model under a harness teams had already standardized on, and nobody had been assigned to respond. In markets, an AI agent now places real orders through a broker, and an order can be smuggled in by text the agent merely read. On the technostress side, the research found people losing the thread of which ideas were even theirs. The agent era did not arrive as a capability gap this week. It arrived as an accountability gap, and it is wide open at every layer.
What we published
AI adoption this week
The Q3 Series A question shifted from what gross margin will be at scale to what it survives at right now, under agent-driven load.
Anthropic split agent billing into its own credit pool, so I argue the live Series C question is who owns the agent cost ledger before the board meets.
Two agent-identity surveys showed nobody can name every agent in production; the fix is a one-page register with a named owner, not a vendor RFP.
With Meta cutting 8,000 jobs and a NY Fed paper finding no AI-driven decline in postings, I separate the two mechanisms a CEO is choosing between.
Gartner's $2.59 trillion 2026 AI forecast pairs with cautious enterprises, so I read the three signals as money buying control, not transformation.
Most products labeled agentic are repackaged chatbots, so I give a Series A founder a one-week way to vet a vendor properly.
The EU published high-risk AI classification guidelines the same week the White House postponed its AI order; here are five decisions that hold up either way.
AI coding agents this week
Microsoft's MDASH harness surfaced 16 production Windows CVEs in a week, and the same harness shape sits on every engineer's laptop.
Microsoft revoked Claude Code from the org behind Windows even though developers loved it, which ends the bottoms-up harness adoption story.
GitHub flipped the Copilot base model and OpenAI merged ChatGPT and Codex the same weekend; the question is who, by name, owns the response.
In 72 hours Microsoft swapped Copilot's model, Cursor shipped a model at a tenth of frontier price, and Google brought a free agent-first IDE to I/O.
With two model launches inside a week, I show how to build a harness evaluation that survives the release cadence instead of expiring with it.
Falling model prices lift the harness ROI number with no extra output, so the fix is reading the ratio as two numbers, cost and output.
GitHub's own repositories were breached through a VS Code extension, and I give the three questions a board will ask after it.
AI and markets this week
AI quant strategies trailed discretionary managers by six points in April, and I trace the gap to model convergence, not classical signal decay.
An AI agent's stock order passes through two independent routing decisions, and the SEC's Rule 605 disclosures audit only one of them.
Ackman had exited the Alphabet stake the public 13F still disclosed; the 45-day staleness window is the whole mechanism worth understanding.
With the Fed minutes and the Nvidia call the same afternoon, synchronized AI reading has moved the news-trading edge from speed to tone.
Exchanges broadcast a live order-imbalance feed for the final ten minutes before the close, so the day's fairest price is shaped by whoever reads it fastest.
AI agents wired into brokerage accounts read text they cannot fully trust, so the question is what an agent can read and act on without approval.
Alternative-data edges do not fade gently; they drop at a cliff the moment a vendor's exclusivity contract lapses.
Self-awareness in the age of AI this week
A diary study of 173 working adults found the same AI-heavy work can feel like a stretch and a quiet threat on the same day.
A peer-reviewed mindfulness study points at a small stretch of curiosity that does real work before a decision, the kind AI tools tend to skip.
A new study found people could not reliably remember a week later which ideas were theirs and which came from an AI chat window.
A three-wave study of 507 professionals found felt loss of skill or autonomy becomes an identity threat that drifts people from their own work.
Research following 912 people found that whether someone checks an AI output depends on the stance they take toward the tool, not their speed.
A study found people who most dislike boredom feel it more often and reach for screens more, worth a leader's attention as AI fills empty moments.
New research names a task AI drafting quietly adds: the unpaid work of rewriting machine output until it reads as genuinely yours.
Signals to implications
Signal. Two agent-identity surveys this week found that no organization can reliably name every agent running in production, what it touches, or who can shut it off.
Implication. Before Q3 close, produce a one-page agent register with a named human owner per agent. It is a list, not a procurement cycle. [Exec | Founder]
Source: The agent register question Series B boards should ask before Q3 close
Signal. GitHub flipped the Copilot Business base model on May 17 and OpenAI merged ChatGPT and Codex the same weekend, with no engineering team choosing either change.
Implication. Assign one named person to own your response when a vendor swaps the model under your harness. That name matters more than which harness you picked. [Eng Leader]
Source: Who on your engineering team owns the model swap when the harness changes under them?
Signal. When an AI agent places an order through a retail broker that sells its order flow, the order passes through two routing decisions, and Rule 605 disclosures audit only one.
Implication. Before letting an agent trade, ask the broker which of the two routing layers their published execution-quality numbers actually cover. [Investor]
Source: The two routing layers between an AI agent and your fill price
Signal. A week after working with an AI chat window, people in a new study could not reliably tell which ideas were theirs, and the blended ones broke memory most.
Implication. Notice when you are mixing your own thinking and the model's in the same pass, and mark which is which while you work rather than after. [Self-aware Worker]
Source: The AI Memory Gap: When Your Ideas Stop Feeling Like Yours
Signal. Coding-agent vendors cut model prices sharply this week, which lifts the harness ROI ratio on a board slide even when the team ships nothing more.
Implication. Read the ROI ratio as two separate numbers, cost and output, and confirm which one actually moved before you report it. [Exec | Eng Leader]
Source: Your coding agent got cheaper this week. Your team did not get faster.
The contrarian take
The week read like a price war that buyers won. Vendors cut model prices, Cursor shipped a model at a tenth of frontier cost, and the harness ROI line on board slides drifted up on its own. That cheaper number is the trap. As I show in your coding agent got cheaper but your team did not get faster, a falling cost lifts the ratio while the team ships nothing more. The same week, the agent register question showed those orgs still cannot name who owns each agent in production, while the trading-agent piece showed an agent can place an order nobody approved. The scarce resource this week was never compute. It was a named human attached to every agent that acts. Spend the next two weeks building that list, not renegotiating that invoice.
Next week
If this recap was useful, the newsletter delivers it straight to your inbox every Monday. Subscribe here.