AI agents have reached a point of practical utility where they can tackle complex, real-world enterprise challenges like large-scale code migrations and automated cybersecurity, delivering massive efficiency gains.
The US government's reliance on archaic COBOL-based systems for critical functions represents a significant, urgent, and addressable market for AI-powered modernization.
The role of a human software engineer is rapidly evolving from a creator of code to a director of AI systems, focusing on high-level specifications, verification, and objectives.
The competitive landscape of AI favors model-agnostic 'agent labs' that focus on customer outcomes over the foundational model providers themselves, as they can flexibly orchestrate the best combination of tools for a given problem.
Traditional business moats, particularly those built on high customer switching costs, are becoming obsolete in an era where AI can automate complex migration and integration tasks.
2010s
Cites a GAO study from this period that identified 10 critical legacy government systems needing modernization, noting that only 3 have even started the process, highlighting the problem of government inertia.
2024-03
Cognition officially launched its AI software engineer, Devin.
Approx. 2024-06
States that after approximately 3 months of internal use, Devin became the number one contributor to its own codebase at Cognition.
2025
States that Cognition acquired the agentic IDE Windsurf.
By 2028
Predicts that human engineers will primarily write high-level specifications in English rather than writing production code.
▶AI as a Modernization Catalyst for Government and EnterpriseMay 2026
Kaplan repeatedly emphasizes the immense challenge and opportunity presented by legacy IT systems, particularly the tens of millions of lines of COBOL powering critical US government functions. He positions AI agents as the key technology to finally overcome the inertia that has stalled modernization for decades, citing the US government's $100B+ annual IT spend as a massive addressable market.
This theme suggests that the most valuable initial markets for advanced AI agents may not be in creating novel applications, but in solving decades-old, high-cost legacy system problems that were previously considered intractable.
▶The Automation of Software EngineeringMay 2026
Kaplan outlines a clear trajectory for the role of the software engineer, moving from a hands-on coder to a high-level architect who writes specifications in natural language. He supports this with evidence of AI performance, such as Devin becoming a top contributor to its own codebase, and makes a firm prediction that this shift will be mainstream by 2028.
Investors and analysts should monitor the development of tools like 'Devin Review,' as the value chain in software development is shifting from code creation to code verification and system design, creating new market opportunities.
▶The AI-Driven Cybersecurity Arms RaceMay 2026
Kaplan presents a dual-use case for AI in cybersecurity. On one hand, he details how malicious actors are using AI to find novel vulnerabilities. On the other, he highlights how large financial firms use Devin to automatically triage and remediate CVEs with a 70% success rate, creating a high-speed, automated defense against these new threats.
This dynamic creates a compelling, non-discretionary demand for defensive AI agents, as organizations cannot afford to fall behind in an arms race where the speed of automated attack and defense is the deciding factor.
▶Cognition's Model-Agnostic Agent Strategy
Kaplan argues that 'agent labs' like Cognition hold a structural advantage over foundational model providers. By remaining model-agnostic and focusing on customer outcomes, they can select the best tool for any given task, as evidenced by internal research showing multi-model agents outperform single-model ones. This is complemented by a go-to-market strategy that includes high-value government contracts, enterprise sales, and community building via free tools.
Cognition's strategy positions it as a value-added layer on top of the foundational model 'stack,' suggesting the long-term winners in the AI space may be those who master application and workflow, not just raw model performance.