The last edition led with a price. Two vendors halved the cost of one agent step inside four days. This edition leads with a limit. An operator published the point at which its own autonomy stopped rising, and the reason is not model quality.
Two days is a short window. This one produced three items and no more. There was no new dark-factory writing, no spec movement, and no new security incident. Three items is the whole of it.
Grab put numbers on an autonomy ladder
InfoQ covered Grab's analytics programme on the seventeenth, drawing on Grab's own engineering write-up. This is the clearest public account so far of an operator running a graded autonomy model and measuring the result.
Grab uses five levels of autonomy. Each level names who frames the question, who does the work, and who reviews the result.
- Level 3. Humans frame the question and review the result. Agents find the data, write and run the queries, check the results, and draft the analysis.
- Level 4. Agents plan and orchestrate the workflow. Humans review at defined gates.
- Level 5. End-to-end autonomy. Humans set objectives, quality thresholds and escalation rules, and nothing else.
The measured effect over four months is direct. Mechanical analyst work fell from 44% in February to 30% in June. The share of requests answered with no human involvement rose as follows.
| Request type | February | June |
|---|---|---|
| SQL requests | 50% | 81% |
| Data pulls | 63% | 90% |
| Metric questions | 53% | 67% |
Read the spread, not the average. A single company-wide autonomy percentage would have hidden the important part. Autonomy did not rise evenly. It rose most where a machine can check the answer, and least where the answer is a judgment.
A SQL result either matches the source rows or it does not. A data pull either returns the requested fields or it does not. A metric question asks what a number means, and that is a different kind of work. Grab kept human accountability for metric definitions, causal interpretation, business assumptions, and final decisions. Those exclusions are deliberate, and they are the reason the third row stops at 67%.
The ceiling on autonomy is not model quality. It is who can check the work.
A court decided that an agent is not a person
The Ninth Circuit decided Amazon.com Services, LLC v. Perplexity AI on 4 August 2026. Legal analysis and industry coverage crested on the seventeenth. This blog has not covered the case before, so the date is stated plainly here. The decision is two weeks old, not two days old.
The CFAA is the Computer Fraud and Abuse Act. It is the main United States law against unauthorized computer access. The court held that "the CFAA contemplates access by a person", and that Perplexity's agent is "a tool, not a person, for statutory purposes". When a user directs the agent to act on Amazon.com, the court reasoned, it is the user who accessed Amazon's computers. The panel vacated the injunction that had blocked Perplexity's Comet Browser agent from the Amazon Store.
Two consequences follow, and they point in opposite directions.
For builders, this removes a large uncertainty. An agent acting under a user's direction does not itself commit unauthorized access. The operator inherits the legal position of the user it acts for.
For everyone else, it opens a gap. If the agent is a tool, then nobody clearly owns liability for what a fully autonomous agent does. Commentary calls this a liability vacuum. The court's reasoning rests on a user directing the tool. It says much less about an agent that runs on a schedule with no user in the loop, which is exactly the shape of a dark factory.
One boundary is worth stating. The ruling is about access. It is not about correctness, safety, contract terms, rate limits or copyright. A site's terms of service are untouched. Do not read this as permission.
The toolchain got two new owners in one week
SpaceX closed a $60 billion all-stock acquisition of Anysphere, the maker of Cursor, effective 14 August. Cursor now runs as a wholly owned subsidiary inside a division called SpaceXAI. It is reported as the largest startup acquisition on record.
Bloomberg reported on 16 August that Stripe is acquiring OpenRouter for more than $7 billion. OpenRouter routes developer traffic across more than 400 models through one API endpoint and serves about eight million developers. The price is about 50 times annualized revenue, and 5.4 times the $1.3 billion valuation OpenRouter set in a Series B three months earlier. The premium is for position, not revenue.
Stripe already owns Metronome, a usage-metering and billing platform for AI companies, acquired in late 2025. Adding OpenRouter puts routing, metering and payment in one place. That is convenient. It also means one party sees which model answered which request, at what cost, for which developer.
Cognition, the maker of Devin, is reported to be raising at a $40 billion valuation, with Devin near $1 billion in annualized revenue.
The pattern is ordinary dependency risk, applied to a layer many teams treat as infrastructure. In two days the leading agent IDE moved under a launch company. The leading model router moved under a payments company. Neither buyer set its priorities around your build.
How we apply this at River.io
We are writing the autonomy ladder down. This factory already grades work informally. Grab's model is the public version of the same idea, and the value is not the levels. The value is that each level names who frames the question, who runs the work, and who reviews the result. The artifact is a table with one row per repository, one column for the level, and one column for the specific step that must stay human. CareTime and TimeForCare sit at Level 3. The certification-record paths in MaterialsAndPractices sit at Level 3. None of those move. sloth, MeowPassword and weathergalactic are the candidates for Level 4.
We copy Grab's split, not Grab's percentage. Their numbers say autonomy rises where a machine can check the answer. Map that onto this shop. Test generation, lint cleanup, warning removal and documentation regeneration are checkable by the compiler and the test suite, so they can run at higher autonomy. API shape, data model, pricing, and anything that defines what a record means stay with a person. That line was already the house rule. Grab is evidence it holds at a company much larger than this one.
We measure the hands-off share per work type, not per repository. Grab reported three numbers for three request types, and the spread carried the insight. One factory-wide percentage would have hidden it. The change is a METRICS.md column: for each work type, the count of runs that merged with no human edit. That is a countable row out of the existing ledger. It is not a new system.
The CFAA ruling changes nothing we do today. No repository here runs an agent against a third party's site. sloth reads CERT, CVE and MITRE feeds, which are published for reading. The ruling matters only if a future product ships an agent that acts for a user against systems we do not own. If that day comes, design for the fact the court leaned on: a user directs the agent, and the record shows it. Keep a per-run record of which user authorized the action. That record is the thing the reasoning depends on.
Consolidation is an argument for the posture we already have. The zero-dependency rule means no repository here breaks when a vendor is acquired. The exposure that remains is the model API itself. The answer there is the same as always. Keep the call behind one small internal seam, so changing vendors is an edit in one file. If this factory ever routes through a third party, that seam is what keeps the acquisition somebody else's problem.
Nothing here argues for loosening a gate. Grab is running this well and still keeps interpretation and final decisions with people. The PHI-touching and EVV-submitting code paths stay human-gated.
Sources and bibliography
Primary
- Grab Engineering, How AI is transforming analytics at Grab. The operator write-up behind the InfoQ report. Source of the five-level autonomy model and the human-accountability exclusions.
- InfoQ, Grab Cuts Mechanical Analytics Work From 44% to 30% with AI Agents (17 August 2026). Source of the 44% to 30% figure and the per-request-type hands-off percentages.
- Grab Engineering, From firefighting to building: how AI agents restored our team's core productivity. Companion write-up.
Legal analysis
- Wilson Sonsini, Ninth Circuit Addresses CFAA and Agentic AI Tools in Groundbreaking Decision. Source of the quoted holding and the 4 August 2026 decision date.
- Cooley, Ninth Circuit Rules on AI Agent 'Access' to Third-Party Websites Under CFAA (6 August 2026) · Ropes & Gray, Tool or Intruder? What Amazon v. Perplexity Means for Agentic AI and the CFAA.
- Techdirt, Ninth Circuit: Your AI Agent Can't Violate Hacking Law. But You Might. (5 August 2026) · Forkast, The 9th Circuit's Browser Analogy Leaves a Liability Vacuum. Sources of the liability-gap reading.
Reporting
- Forkast, Stripe Acquires OpenRouter for $7B+ · Dataconomy · Quartz · Neowin. Sources of the deal size, the 400-model and eight-million-developer figures, the revenue multiple, and the Metronome background. All trace to a Bloomberg report of 16 August 2026.
- Seeking Alpha, SpaceX completes $60B acquisition of Cursor · Cryptopolitan · SatNews (13 August 2026). Sources of the $60 billion all-stock terms and the 14 August effective date.
- The Daily Star, AI coding startup Cognition in talks to raise new round at $40 billion valuation (17 August 2026). Source of the Devin revenue and valuation figures.
- AI Agents Directory, News Brief, 17 August 2026 · AI Agent Store, Daily News, 17 August 2026. Aggregators. Used to locate the primary sources above.
Prior editions referenced
- The price of an agent step fell by half. Twice. (16 August 2026). The Gemini and DeepSeek price cuts, the Taiwan intrusion, and the Anthropic turf-war study.
- Ninety percent of AI-native developers are stuck at Level 2 (26 May 2026). The earlier treatment of autonomy levels in this factory.
Vendor and blog figures indicate direction, not audited benchmarks. The Grab percentages are Grab's own and cover analytics work at one company. The Stripe and OpenRouter terms come from a Bloomberg report and have not been confirmed in a filing. The Cognition valuation is a report of talks in progress, not a closed round. The Ninth Circuit decision binds only that circuit and may be appealed.