An Anthropic staff member recently observed a machine struggle with one of the hardest problems in mathematics. It was the Riemann hypothesis, dating back to 1859, and its complete solution has a million-dollar bounty. The staffer wasn't a mathematician, but he used variations on phrases like "keep going" or "believe in yourself."
After hundreds of failed attempts where the agent doubted itself and stopped, he prompted the agent to try again. And again. Which finally led it to achieve breakthroughs humans haven't been able to accomplish in this area of mathematics, all without any of the advanced technical background previously needed to do this work.
The work was checked by Anthropic's own mathematicians, outside number theorists, and a formal proof checker.1 In another study, Anthropic also found encouragement also helped Claude disprove the Jacobian conjecture.
Other researchers have measured similar effects. In March, a single-author preprint reported a controlled experiment on Claude Sonnet 4. Agents given a trusting, supportive framing caught 59 percent more hidden bugs than agents given none, and took 74 to 83 percent more investigative steps to get there, while fear-based pressure improved nothing.2
OpenAI reports that simply telling an agent to persist through uncertainty raised its score on a hard software benchmark by close to 20 percent.3 Microsoft researchers found the same effect in 2023.
Anecdotally, many engineers including those at Codespeed have found that positioning agents to question prior assumptions from first principles, consider if there are more efficient solutions than the current path, and to scrutinize its own work also produce stronger results. These workflows prove that prompting a mixture of encouragement and regular questioning of oneself pushes agents to achieve more valuable solutions and better quality code.
As software engineering evolves, it's clear that writing code is far less of the essential equation. Effective communication, management and strong leadership in both business and tech are taking over instead.
The engineers who bring a clear vision, communicate precisely, and manage their agentic teams in the same ways the highest output human teams have operated at top companies for decades will continue to see incredible compounding results.
In 2026 and beyond, AI software development is far more than a set of tools and features. It's a force multiplier for the best leaders on your team.
