Instagram’s co-founder stepped down as Anthropic’s CPO to code hands-on inside Anthropic Labs.
Comparing his Sept 2025 forecast with how they actually ship in 2026 exposes the whole playbook:
Overnight agents, FFmpeg QA loops, and the 10:1 pivot rule. https://t.co/uoSWZdGLgN
@adityaa_behera gemini 5 next? where is a gemini krypton , phyton, Oganesson, pluton, uranium, helium? everyday i reading news which models gemini are next😂
Appreciate it. Honestly the wildest part to me is that even Krieger admitted he “feels like a total newbie again” — six months was enough to make 2025 workflows feel outdated.
Feels like taste and knowing what to cut are about to separate good builders from everyone drowning in generated code.
In 2010, Mike Krieger built Instagram v1 in five days of all-nighters.
Hashtags alone took a full week.
In January 2026, he stepped down as Anthropic’s CPO to co-lead Anthropic Labs and build hands-on again.
Five months later, he kicked off a personal mobile + web app on a Saturday morning, went hiking with his kids, steered edits from his phone, and came home to a working product.
I went back through his Sept 2025 forecast and his June 2026 Anthropic Labs interview side by side. Five takeaways completely reframed how I think about building with AI:
1. “I feel like a total newbie again.”
Even the guy helping build Claude had to throw out his 2025 prompting habits. Step-by-step instructions are dead — now he states the intent, lets agents run overnight, and wakes up to finished work (including self-built mock backends when a remote service goes down).
2. Getting “verification-pilled”
When agents ship thousands of lines while you sleep, manually reading diffs stops working. Anthropic Labs relies on PR screenshot galleries, FFmpeg video scrubbing so Claude spots animation jank itself, in-memory Postgres/Redis mocks, and Slack MCP threads where “Mike’s Claude” ships fixes and follows up.
3. Classic chat UI is breaking
Krieger doesn’t read walls of text — he asks Claude to generate interactive HTML pages with diagrams. He runs 5–6 parallel tabs on remote dev boxes, routes trivial questions like NBA Finals stats back to Sonnet, and says teams now need multiplayer agent workflows.
4. The 10:1 pivot rule (earned the hard way)
Shutting down Artifact taught him a brutal filter: 1 unit of effort returning 10 units of pull means push. 10 units of effort for 1 unit of lukewarm “it’s okay” means exit.
5. What AI still can’t replace
Anthropic’s Project Vend proved autonomy without business judgment just makes mistakes faster. Taste, point of view, and production judgment (knowing when to just restart a leaking server on a weekend instead of refactoring) remain 100% human.
Code stopped being the bottleneck. Verification and clarity took its place.
I broke down the full 14-month shift + 4 visual diagrams and his 6 daily habits in the article below.
Which part of this workflow are you already using — and what still feels impossible in your stack?
Instagram’s co-founder stepped down as Anthropic’s CPO to code hands-on inside Anthropic Labs.
Comparing his Sept 2025 forecast with how they actually ship in 2026 exposes the whole playbook:
Overnight agents, FFmpeg QA loops, and the 10:1 pivot rule. https://t.co/uoSWZdGLgN
Right? Going from 5 all-nighters for Instagram v1 to long-pressing a button on a trail to dispatch edits and preview them live on your phone is wild.
Feels like once you switch to mobile + remote dev boxes, going back to babysitting one laptop terminal feels ancient. Are you already running agents remotely or still mostly local?
@0numex Record profits, best year ever, CEO at $90.6M (421x the median), and still handing 60k people a 'voluntary' exit form. The 'vote with your feet' line while saying 'I don’t want you to leave' is peak corporate. Everyone got the real message