Lovable is merging 150+ PRs a week with one human per swarm
Lovable published a candid write-up this week on scaling agentic coding, and the numbers are the story. In January they were merging 20–30 pull requests a week with plan mode, manual permission prompts, and normal code review. By June: 150+ merged PRs a week, with one human overseeing six or seven agents, each running its own swarm of subagents. Total token spend since January is about $85,000, running around $25k a month now.
The interesting part isn't the velocity, it's what they had to build to survive it. They classify every PR by risk and hard-enforce human review on the high-risk ones — you can't eyeball 150 PRs a week, so you triage. Code review shifts from reading lines to validating specs and gating by blast radius. That's the real transition: humans moving from writing code to designing the system that supervises the code being written. (This is the exact scenario I wrote about in the Sonnet 5 piece — that kind of velocity is only safe if your architecture has hard boundaries the agents can't cross.)
Gartner put a number on the SaaS threat: $234 billion
On July 1, Gartner warned that up to $234 billion in enterprise application spend is exposed to what they're calling "agentic arbitrage" by 2030 — roughly a fifth of enterprise SaaS. The mechanism is straightforward: agents do work across systems and quietly bypass the UI-heavy apps you pay for per seat. Agents don't log into fifty dashboards; they call APIs.
I think this is directionally right. If you sell SaaS, your moat stops being your interface and becomes whatever outcome you can actually guarantee. If you buy it, there's a real opening to build an orchestration layer that sits across your vendors instead of paying for a seat in each one — though the thing you'd build to do that is itself an agentic system, so it needs the same architecture discipline or it becomes its own mess.
AI agents are now making real card payments
CaixaBank and Visa completed a real purchase initiated by an AI agent on live card rails — tokenization, identity checks, fraud monitoring, an actual merchant. BBVA and others are close behind, and Visa is rolling this out through its agentic-commerce program across Europe. Agents just went from read-only assistants to authorized buyers.
If you're building anything that will eventually spend money or reconcile it, design for this now. Consent, spending limits, and an audit trail aren't features you add after the demo — they're the difference between an agent that can transact and one you'd actually let transact.
Oh, and b/c of this Government of Singapore has a framework for safe agentic payments: https://www.mas.gov.sg/publications/monographs-or-information-paper/2026/safeguards-for-agentic-finance-at-runtime
China starts regulating agents by kind
China's Interim Measures for anthropomorphic AI interaction services take effect July 15, aimed at emotional and companion bots: algorithm filing, AI-disclosure at login and every two hours, and hard rules around minors. Workplace productivity agents are largely spared, as long as they don't simulate a personality you're meant to bond with.
That's a useful signal even if you never ship in China. The regulatory heat is pointing at personified, emotionally sticky agents, and Western bills aimed at kids and companion chatbots point the same way. If you're building serious tools, keep your agents cold and outcome-focused. It's the better product instinct anyway.
Frontier coding is bumping into export controls
For about three weeks in June, the U.S. government suspended foreign-national access to Anthropic's most capable models, Fable 5 and Mythos 5, after Amazon researchers showed Fable could be prompted into finding software vulnerabilities and, in one case, writing code to exploit them. Commerce lifted the order on June 30 once Anthropic committed to proactively detect and report those risks.
Set the specifics aside and look at the shape: frontier coding capability is now close enough to offensive-security capability that export controls are in the conversation. If your work touches automated code review or security tooling, expect more guardrails on the models you depend on — and plan for the occasional refusal that shows up as a silent failure rather than a clear error.
Of the five, the Lovable numbers are the ones I'd stare at longest. That's not a lab result or an analyst's forecast — it's a shipping team telling you what a swarm actually costs and what you have to build to keep it from hurting you.
Jason Vertrees is the founder of Heavy Chain Engineering (heavychain.org), which helps lower middle-market vertical SaaS companies and PE firms turn scattered AI usage into measurable delivery leverage — 85% faster feature velocity, six-to-eight-week projects shipped in days. If you want help building an AI-native engineering organization, book an AI Delivery Assessment (app.onecal.io/b/jason-vertrees/30-mins).