.png)
Goldman Sachs
Finance · US
40% faster code delivery
Goldman Sachs deploys an autonomous AI coding agent alongside 12,000 developers, calling it a hybrid workforce
The first major bank to deploy a fully autonomous AI coding agent in a production pilot, not a suggestion engine but an agent that completes coding tasks end-to-end. The CIO frames this publicly as a hybrid workforce of 12,000 people and AI agents.
Friction
Goldman's 12,000 software developers spent significant time on routine work including unit tests, documentation, boilerplate code, and legacy modernisation. This slowed strategic innovation and allowed technical debt to accumulate.
Breakthrough
A dual deployment: a GitHub Copilot-variant GenAI coding assistant trained on Goldman's internal codebase, plus an autonomous Devin agent that operates in a sandboxed compute environment, independently planning, writing, testing, and debugging code before a human reviews and merges.
Impact
40% faster delivery of standard coding tasks. 15% fewer post-release bugs. 25% faster onboarding for new developers. 70%+ daily adoption rate. Expected productivity gain: 20%, equivalent to the output of 14,400 from 12,000 people. All with full audit logging for regulatory compliance.
Unlock the full analysis with breakthrough, impact, what made it smart and its technical approach below!
Problem
Investment bank software is among the most complex and compliance-sensitive in the world. Getting AI-assisted coding deployed at Goldman required solving the same trust problems as any other AI in regulated finance: explainability, audit trails, and clear human oversight at every decision point.
What made it smart
The distinction between Copilot, which offers suggestions the developer accepts or rejects, and Devin, which executes autonomously in a sandbox, represents a fundamental step change in how AI integrates with knowledge work. Goldman's willingness to pilot genuine autonomy rather than just augmentation is the leading edge of enterprise AI deployment.
Technical approach
GenAI coding assistant trained on Goldman's internal codebase and project documentation provides real-time code suggestions within the developer's IDE. Devin runs in a sandboxed environment with access to a defined code repository, able to plan tasks, write code, run tests, and iterate through debugging cycles independently. Humans review output and merge.
Strategic lesson
The distinction between AI augmentation and AI automation in knowledge work is not a matter of technology. It is a matter of how clearly you can define the task, constrain the environment, and design the human review checkpoint.
Reflection question
In your engineering or knowledge work, which categories of tasks are well-defined enough that an AI agent could execute them end-to-end, if given a clear scope, a constrained environment, and a human review before implementation?