Poolside, founded in 2023 by former GitHub CTO Jason Warner and Eiso Kant, builds the Laguna family of open-weight foundation models specifically for agentic software engineering. The company targets enterprises with over 5,000 developers, including banks and defense contractors, through sales-quoted contracts that bundle per-developer seats with dedicated compute resources. Since July 2026, Poolside has also distributed its models via third-party marketplaces, though it shut down its own hosted inference platform on October 31, 2026. This shift forces large organizations to self-host the weights on their own infrastructure, a move that aligns with the strict data residency requirements of financial institutions like Goldman Sachs.
The strategic pivot from a shared SaaS API to private deployments reflects the reality that frontier AI coding tools now require deep integration into existing security frameworks. For firms managing sensitive trading algorithms or customer data, running inference inside a private VPC or on-prem is non-negotiable. Poolside’s business model embeds forward-deployed research engineers directly with customer teams, ensuring the models adapt to specific legacy codebases rather than just generating generic snippets. This approach contrasts sharply with consumer-facing tools that rely on public cloud endpoints and standardized prompts.
Goldman Sachs: From test generation to autonomous agents
Goldman Sachs has integrated multiple AI systems into its engineering workflow, starting with GitHub Copilot for all 12,000 developers and expanding to Cognition’s Devin for autonomous coding tasks. Chief Information Officer Marco Argenti describes Devin as a "new employee" capable of scoping, writing, testing, and debugging production code, estimating it is three to four times as productive as previous AI assistants. The bank reported a 20% average increase in developer productivity, translating directly into cost savings across its technology division.
The quality assurance team faced a distinct challenge when deploying these agents: how to validate code when the same AI generates both the implementation and the tests. To address this, Goldman uses Diffblue Cover for automated Java unit-test generation, which increased code coverage from 36% to 72% in one backend module within 24 hours, a task that would have taken eight days manually. In another component with 15,000 lines of code, the tool generated over 3,000 unit tests overnight, roughly 180 times faster than manual effort. However, Argenti notes that high test-pass rates do not guarantee correctness if the agent makes mistaken assumptions about ambiguous requirements.
The bank now employs a multi-model strategy, using GS AI to select between GPT, Gemini, and Claude models based on specific task accuracy and cost efficiency. Agentic models produce ten times more tokens than standard reasoning models, making them expensive but necessary for complex research tasks like analyzing comparable companies. Human approval remains essential for changes involving critical applications or regulated decisions, ensuring that AI speed does not outpace regulatory compliance.
Stripe: Prototyping at the frontier
Stripe’s AI Build team focuses on building with leading AI companies, requiring solutions engineers to prototype rapidly using modern coding harnesses like Cursor, Claude Code, or Codex. The role demands navigating incomplete products and interoperability gaps, turning customer needs into reusable reference architectures and implementation playbooks. Engineers must work across multiple Stripe products to design solutions that span billing, payments, and revenue management.
The job description explicitly lists proficiency in Python, JavaScript/TypeScript, Java, or Ruby, emphasizing the ability to learn quickly and build credible prototypes in ambiguous environments. Stripe operates at startup speed, expecting staff to lead through influence across sales, product, and engineering teams while maintaining clear ownership of customer outcomes. This structure ensures that field learning feeds directly into the product roadmap, closing the gap between emerging AI capabilities and stable financial infrastructure.
Unlike Goldman’s focus on validating AI-generated code against existing standards, Stripe’s approach centers on creating new patterns for AI-native workflows. The team coordinates account teams and professional services to activate proof-of-concept engagements, ensuring that successful customer work becomes repeatable solution components. This method allows Stripe to scale its financial services to the fastest-growing AI companies without waiting for traditional enterprise procurement cycles.
