Conference Presentation, Fireside Chat, Interview
The Future of Software Creation with Replit CEO Amjad Masad
- Software engineering is projected to transition from an expert-only domain to one accessible to anyone, with software engineering agents expected to achieve 70-80% on the SWE Bench benchmark as models improve significantly within two months.
- Agent autonomy levels are forecast to progress from the current "Level 3" (10-15 minute operation with human input) to "Level 4" (fully autonomous with some attention required) and eventually "Level 5" reliability within the next couple of years, enabling the simultaneous deployment of thousands of agents with 95% success rates.
- Computer use capabilities for models are predicted to see major improvements over the next three to six months, potentially automating many real jobs, while sampling and simulations aim to increase agent reliability by two to three folds.
- Universal model access will soon allow direct app integration with automated billing, and future agents are expected to possess financial wallets for service payments and capabilities to hire humans or other agents via platforms like TaskRabbit or agent markets.
- The value of all application software is anticipated to approach zero over the order of years, shifting the replaceability of generic SaaS from 15% to 100% within the next few years, forcing a business pivot from creating applications to solving problems.
- Corporate structures are expected to evolve from hierarchies to network-based organizations resembling open source projects, while job roles will become less specialized as generalist employees capable of generating broad business value replace highly siloed professionals.
- Transaction costs for hiring developers or agents are predicted to drop to zero, facilitating hiring with a single click, while the most critical asset for building a startup agent will be domain knowledge rather than general coding ability.
- Training methodologies will shift from reliance on human code to AlphaZero-style self-play and reinforcement learning to overcome error accumulation plateaus in agent-to-agent code generation and improve long-term reliability through parallel trial and error.
- Human ingenuity will remain essential for solving truly novel problems outside of distribution, making liberal arts and critical thinking more valuable in education to prepare a workforce with a broad worldview, with early startup involvement recommended for acquiring generalist experience.