- Stripe thinks we achieved singularity (i.e. the moment AI bots started running the economy) at the beginning of this year https://www.youtube.com/watch?v=e13-s0p1tfE. Some learnings/highlights from their recent conference:
- Processed $1.9 trillion in payment volume last year, 1.6% of global GDP, growing 34% year-over-year while the global economy grew 2%. They have a live GDP ticker on their site now https://stripe.com
- 99.9996% API availability while shipping over 1,000 code changes a day. Infrastructure for 86% of the Forbes AI50, 90% of the Dow, and 80 companies in the NASDAQ 100
- The 2026 Atlas cohort is tracking 5x last year’s revenue. Around 5 million US solopreneurs now make over $100k. New companies launch globally on day one. AI has collapsed the cost of discovery and coordination, so firms can stay small and let the market do more of the work
- AI agents are no longer just chatting. They are buying, negotiating APIs, and executing tasks. Stripe launched a “wallet for agents” and a blockchain called Tempo so AI can stream fractions of a cent in stablecoins per API call. Commerce is moving from a single checkout moment to a spending policy where humans set the guardrails and agents execute within them
- For other interviews, Nat Friedman calls this the “slow phase” because humans are still in the loop. Once models self-improve, the curve goes vertical. John Collison brings up the Solow paradox: electricity didn’t show up in productivity stats until factories were redesigned around it, and we’re in that replatforming period now. Sam Altman thinks the bigger story isn’t software, it’s what cheap intelligence does to material science and biology
- They’ve also made their full product roadmap public, which is kinda cool and very inspiring. 200+ products and features in active development https://stripe.com/roadmap
- Coding agents are powerful because they remove friction from making software, but some of that friction was useful. It gave us time to think, feel the cost of complexity, notice bad abstractions, and understand the system as we built it. When agents generate code too quickly, small errors, duplicated logic, and shallow architecture can compound faster than humans can review or absorb them. The lesson is not to avoid agents, but to slow the loop down: use them where the task is scoped and feedback is clear, while keeping humans close to the architecture, APIs, and product decisions that define the shape of the system. AI can help write code, but humans still need to preserve agency over what is being built and why https://mariozechner.at/posts/2026-03-25-thoughts-on-slowing-the-fuck-down, https://www.youtube.com/watch?v=_Zcw_sVF6hU
- Symphony, an agent orchestrator built by OpenAI that turns a project-management board like Linear into a control plane for coding agents, shows that agentic programming is bigger than a coding agent: it is the harness that turns human intent into software work by making tasks, tools, workspaces, dependencies, feedback, CI, and review legible enough for agents to execute continuously while humans focus on judgment instead of micromanagement https://openai.com/index/open-source-codex-orchestration-symphony/
- Basecamp’s AI rule is: use AI freely to explore, prototype, and find the right product answer, but don’t confuse a convincing prototype with production-ready code. The core codebase that ships to customers stays protected by human engineering judgment https://public.3.basecamp.com/p/s5B42kpUmSrojy9JPRa45jgn
- Containers are great for packaging software, but they are not where trust should end. They limit what an app can see or use, but everything still shares the host’s Linux kernel, so a kernel bug, risky permission, or mounted host socket can break the boundary. A microVM is a tiny virtual machine (like Firecracker or Cloud Hypervisor) for one workload: it keeps the container-like workflow, but adds its own guest kernel and a stronger hardware-backed isolation layer. For untrusted code, AI agents, CI jobs, and multi-tenant systems, use containers for convenience and microVMs for isolation https://emirb.github.io/blog/microvm-2026/
What I’ve been reading (W18, 2026)
Stripe at GDP scale, slowing the agent loop down, harnessing coding agents, Basecamp's AI rule, and microVMs for isolation