WritingNovember 6, 2026
Uber's Agents Booked 4 Rides. OpenAI's Drafted 948 Bug Patches.

Uber's own CFO said it on an earnings call this week: AI agents booked four rides and five hundred Eats orders out of a combined 320 million trips and orders. OpenAI announced this week that its agents found, verified, and drafted patches for 948 real privilege-escalation bugs in the Linux kernel during an October scan. Same word — "agent." Opposite outcomes.
That gap is the whole digest this week. If you've been pitched "AI agents" for your business — checkout, support, sales — the question isn't whether agents work. It's whether the job you're handing one looks more like Uber's (messy, consumer-facing, low stakes per transaction) or OpenAI's (scoped, verifiable, high volume, boring). This week's news sorted cleanly into those two buckets, plus a third: the money-and-cost conversation every operator should be having right now, because two of the companies minting the most cash in tech just told you how hard they're rationing AI spend.
Where agents actually worked this week
OpenAI's autonomous pipeline found, verified, and drafted patches for 948 Linux-kernel bugs. [1] Not demoed, not just "found" — found, verified, and turned into draft patches, at a scale no human security team was ever going to clear. The job was narrow: a specific bug class, a specific codebase, a specific authorized scope. That's the pattern every agent win this week shares — tight scope, checkable output, no judgment call about whether to bother a human first.
Source: OpenAI — A thousand fewer paths to root
Box's Nemotron-powered retrieval system cuts agent token usage 90%, speeds queries 4x, and drops errors 20%. [2] If your agent bill is dominated by stuffing context into every call, Box's chunking approach — built on Nvidia's Nemotron embeddings, rolling out to Enterprise Advance customers in early 2027 — says the fix is retrieval architecture, not a cheaper model. Translation: before you downgrade your model to save money, check whether you're even feeding it the right context in the first place.
Stripe opened your own revenue data to agents via MCP, and Claude shipped a Plaid-linked Personal Finance tool. [3][4] Ask Stripe's new integration for an MRR-attribution breakdown and you get the underlying SQL back, inspectable and editable, free through January 2027. Claude's version reads your bank and card transactions and flags duplicate charges and creeping fees — Pro/Max/Ultra, US-only beta this week, no ads mixed into the answer. Financial data is turning out to be exactly the kind of scoped, structured, high-stakes-but-checkable job agents are actually good at. Not a coincidence that it's also the job neither vendor is selling as "autonomous."
Where they didn't
Uber's agents booked four rides. [5] Out of 320 million trips and orders, during the company's busiest week in October. The agentic-commerce future every checkout vendor has been selling for a year isn't showing up in the usage numbers of the company best positioned to benefit from it. If a prospect pitches you an "AI agent" for your storefront checkout, ask for Uber's number before you ask for the vendor's demo.
Canva cut 1,271 jobs — 22% of its workforce — and slashed its growth forecast from 30% to 20%. [6] The stated reasons: rising AI compute costs on one side, ChatGPT and Gemini eating its use cases on the other. Design-SaaS is getting squeezed from both directions at once, and "we'll out-feature the chatbots" isn't holding as a strategy for a company that size, let alone a smaller one.
Apple cut engineers' daily AI spending caps from $300 to $200 in October, headed to $100 — and is reportedly weighing a cap as low as $25 a day for non-engineering staff. [7] If the company with the most cash per employee in tech is rationing tokens this hard, your business needs a hard per-seat ceiling before the invoice forces one on you. I wrote about the wrong way to set that ceiling a few days ago — Apple's numbers are the right instinct, a cap, just applied before the bill forces the conversation instead of after.
The hedge: two labs bought their own plumbing
OpenAI acquired Pierre Computer Company, and Meta acquired Conductor — the same week. [8][9] Pierre's Trees and Diffs libraries already power Codex's file navigation and diffing; folding in the vendor that builds your coding agent's internals is how you make sure a competitor can't build on the same primitives. Conductor Cloud shuts down November 20, the local app gets bug fixes "until early 2027," and Cloud subscribers get refunded. Two labs, one week, both buying their coding-agent supply chain instead of renting it. Worth noting if you're betting client infrastructure on any single AI-coding vendor's roadmap: the vendors themselves are hedging by buying their own suppliers.
| This week | The number | What it tells an operator |
|---|---|---|
| Uber's agent bookings | 4 rides, 500 Eats orders / 320M trips+orders | Consumer agentic-commerce isn't here — don't rebuild checkout for it yet |
| OpenAI's kernel-patch run | 948 bugs found, patches drafted | Scoped, verifiable agent work is already production-grade |
| Canva layoffs | 1,271 jobs (22% of staff) | AI compute cost + AI-native competition squeeze SaaS from both sides |
| Apple's AI budget caps | $300 → $200 → $100/day ($25 non-eng, proposed) | Set a token ceiling before the invoice sets one for you |
| Box/Nemotron retrieval | -90% tokens, -20% errors, 4x speed | Agent cost problems are usually a retrieval-architecture problem |
| Pierre + Conductor acquisitions | 2 labs, 1 week | Vendors are hedging their own AI-coding supply chain — you should too |
Here's the shape of it as a flow, because the pattern matters more than any single story:
flowchart TD Job[New agent task] --> Scope{Scope is narrow,<br/>checkable, bounded?} Scope -->|Yes| Win[Works: kernel patches,<br/>financial data, retrieval] Scope -->|No| Flop[Flops: consumer checkout,<br/>open-ended "autonomous" tasks] Win --> Cost{Token cost<br/>dominated by context?} Cost -->|Yes| Retrieval[Fix retrieval architecture<br/>before swapping models] Cost -->|No| Budget[Set a per-seat cap<br/>before the invoice does] Flop --> Wait[Don't rebuild infrastructure<br/>for demand that isn't there yet]
The vendors selling "autonomous agents" for your checkout flow and the vendors buying their own coding-agent plumbing are, this week, the same companies. Watch what they build for themselves, not what they pitch you.
I link that line back to how fractional operators actually scope agentic automation for client work, because the scoping test above — narrow, checkable, bounded — is the same filter I run before recommending an agent over a workflow tool for any engagement. It's also most of what cut one client's software bill 60%: not an agent, a scoping decision.
Sources
[1] OpenAI — A thousand fewer paths to root — openai.com
[2] TechCrunch — Box taps Nvidia's Nemotron for new agent-first content retrieval system — techcrunch.com
[3] Stripe — Analyze your Stripe data directly with your agent — stripe.com
[4] TechCrunch — Claude can now analyze your spending with its new personal finance tool — techcrunch.com
[5] TechCrunch — Almost no one uses AI agents to book their Uber, company says — techcrunch.com
[6] TechCrunch — Canva lays off over 1,200 people amid slowing growth — techcrunch.com
[7] TechCrunch — Apple is reportedly cutting budgets for its internal AI usage — techcrunch.com
[8] OpenAI — OpenAI to acquire Pierre — openai.com
[9] Conductor — Announcement — conductor.build
This is the test I'd run on any "AI agent" pitch landing in your inbox right now: is the job shaped like Uber's or OpenAI's? Narrow, bounded, checkable work is getting genuinely good agent coverage — financial data, code infrastructure, security backlogs. Open-ended, consumer-facing, judgment-heavy work mostly isn't, no matter how the demo looks. Scope the job honestly before you scope the vendor.
The short version
- Uber's agents booked 4 rides + 500 Eats orders out of 320M trips and orders this week — agentic commerce isn't here yet
- OpenAI's agents found, verified, and drafted patches for 948 Linux-kernel bugs in an October scan, announced this week — scoped, verifiable agent work is already production-grade
- Canva cut 1,271 jobs citing AI compute costs and AI-native competition squeezing it from both sides
- Apple capped engineer AI spend at $200/day heading to $100 (considering $25 for non-engineers) — set your own ceiling before the invoice does
- Box's Nemotron-powered retrieval cut agent token usage 90% — agent cost problems are usually a retrieval problem, not a model-choice problem
- OpenAI bought Pierre and Meta bought Conductor the same week — two labs buying their own coding-agent plumbing instead of renting it
Drafted with Claude, reviewed and edited by Bryan before publish.
