The Idea:

In March 2023, Noah Shinn was a Northeastern University undergrad working on a question most of the AI world hadn't got to yet. Chatbots could answer questions. But could an AI agent actually learn from getting things wrong?

His answer was a paper called Reflexion. An agent would fail at a task, write itself a note about what went wrong, then try again with that note in hand. No retraining. No new weights. It was accepted at NeurIPS, one of the field's top conferences, and Shinn dropped out.

He joined Sierra, the AI customer service startup from Bret Taylor and Clay Bavor, as one of its first employees. There, he helped build a benchmark measuring how reliable agents really were. The verdict was grim. The best ones failed more than half of real customer service tasks.

So Shinn knew better than almost anyone how agents break. His bet was that the fix for consumers wasn't another app or dashboard. It was no interface at all. An assistant you could simply text or call, which would then go off and use a phone and a computer the way a person does.

This was the beginning of Instinct.

The Execution:

• March 2023: Shinn published Reflexion as first author, showing agents could improve through written self-reflection instead of retraining. It hit 91% on the HumanEval coding benchmark against GPT-4's 80%, was accepted at NeurIPS 2023, and Shinn left Northeastern that year.
• June 2024: Now one of Sierra's first employees, Shinn co-authored τ-bench. It found GPT-4-based agents completed fewer than 50% of tasks, and only around 25% when repeating the same task eight times. Anthropic later adopted it to benchmark Claude.
• April 2026: Shinn registered Spear Street Technology, the company behind Instinct, in California. Conviction's Sarah Guo and Greenoaks' Neil Mehta led an early round at a valuation above $100M.
• Early August 2026: Instinct opened as an invite-only service, and Kleiner Perkins' Mamoon Hamid led a $75M Series A at a valuation above $500M.
• 24 August 2026: The wobble. TechCrunch reported testers' privacy and security concerns. Instinct had sent an email on Moxxie Ventures founder Katie Jacobs Stanton's behalf without her approval, and tester Claire Vo found 36 Gmail messages still stored hours after revoking access. Its terms granted a perpetual, irrevocable licence to train on user data.
• 26 August 2026: Two days later, Instinct announced a $250M Series B at a $2.5B valuation, co-led by Index Ventures and Benchmark, taking total funding to $350M. The same day, it rewrote its terms to drop the perpetual licence.
• 8 September 2026: Meta launched Muse, its own 24/7 personal agent, which went on to top the US app storeswith millions of downloads. Instinct still has no mobile app.
• 15 September 2026: The Information reported Instinct was in talks at a $10B valuation, with users past 100,000 and the company hitting capacity constraints. Days later, it rolled out a concierge that makes phone calls for you and a "trusted person network" that lets friends' agents coordinate plans.
• Today: Instinct has closed a $1B Series C at a $10B valuation from Sequoia, Benchmark and Coatue, taking total funding past $1.3B. That's 4x its valuation in a month, and 20x in under two months, for a product with no app, no published pricing and no public user numbers.

The lesson? Shinn spent three years cataloguing exactly how AI agents fail, then built one anyway, and it failed in public during its launch week. Investors valued it at $2.5B two days later, then $10B a month after that, even as Meta shipped a rival with millions of downloads. They weren't buying a flawless product. They were buying the founder who wrote the manual on agent failure, wrapped in an interface everyone already knows how to use: a text thread. The best new product might be the one that asks you to install nothing.