The Future of Software Engineering: Cognition’s Russell Kaplan

Cognition's Valuation and Growth Trajectory

• Cognition achieved a $4 billion valuation in March, doubling from $2 billion the prior year and rising from zero at its January 2024 founding.
• The company was founded by a team of former competitive programming Olympiad gold medalists originally as a research lab to solve the "future of software engineering."
• Cognition's flagship product, Devin, is positioned as the first autonomous AI software engineer capable of executing end-to-end tasks, including coding, testing, iteration, and production-ready merging.

Shift from Assistants to Autonomous Agents

• Assistant Era: Defined by synchronous, real-time suggestions (e.g., GitHub Copilot) where humans write code with AI providing line-level completions.
• Agent Era: Defined by asynchronous, autonomous decision-making where humans delegate entire units of work, and agents execute tasks over minutes or hours.
• Agents produce results orders of magnitude more complex than assistant outputs, often delivering fully tested, production-ready code rather than just snippets.
• The transition from "asking questions" to "delegating actions" is identified as a critical game-changer for AI utility and impact.

Operational Capabilities and Real-World Application

• Goldman Sachs CIO Marco Argenti reported writing more code in three months than the previous 15, primarily generated autonomously by Devin.
• Argenti's personal use case included a weekend project where Devin built a full AI recording studio website with custom URLs, knowledge bases, CI/CD pipelines, and model selectors.
• Emerging Roles: The paradigm is shifting toward individual engineers acting as "tech leads" managing a fleet of junior AI agents, requiring the skill to decompose complex projects into delegated units.
• Non-Engineering Use Cases: Devin is being utilized for tasks beyond coding, such as automating lunch catering logistics via DoorDash for an operations lead, leveraging the agent's long-term planning and browser capabilities.
• The artifact produced is primarily code, but the underlying capability is reasoning; this reasoning is expected to expand to other domains where actions can be verified (e.g., policy execution).

Strategic Acquisitions and Product Differentiation

• Cognition recently acquired Windsurf, the maker of the first "agentic IDE," to offer a more human-in-the-loop experience for tasks requiring high-detail creative control.
• Tool Selection Strategy: Cognition differentiates by providing the "right tool for the job": Devin for autonomous execution of defined tasks and Windsurf for collaborative, real-time development.
• The strategy acknowledges that current AI agents act like an "army of junior engineers" requiring steering and scope definition from human supervisors.

Market Focus and Enterprise Adoption

• Cognition deliberately prioritized enterprise clients from inception, finding that integrating AI into existing, proprietary codebases is a harder and more valuable research problem than building greenfield projects.
• The long-term vision predicts a massive increase in the number of people creating software, as the barrier of learning arcane syntax lowers, though current systems still require human intervention for the final 10-20% of complex tasks.
• The "future of software engineering" is defined by higher abstraction levels where users define intent rather than implementation details.

Technological Trends and Performance Metrics

• Capability Doubling: Independent research indicates AI agent capabilities (in terms of task complexity handled) double every seven months, a rate significantly exceeding Moore's Law.
• Model Dependencies: Agent performance closely tracks underlying foundation model releases (e.g., improved performance observed immediately following the release of new model versions).
• Research Focus: Cognition is conducting internal research to adapt text-completion models for the specific task of completing real-world projects and units of work, rather than just predicting the next word.

Future Outlook and Constraints (12-24 Months)

• Growth Trajectory: Domains with formal verifiability, specifically coding and math, are expected to continue exploding in capability due to the ability to automate quality verification loops for reinforcement learning.
• Strategic Goal: Cognition's research target is to promote Devin to the level of a senior engineer within the next 12 months.
• Primary Bottleneck: The leading constraint for continued growth is predicted to be energy and power production, particularly in the United States, rather than data availability.
  • Synthetic data generation has mitigated data scarcity.
  • Compute hunger is increasing; more compute directly correlates with higher capability, creating a potential supply-demand mismatch for power within the next two years.