Tech • AI • Robotics
How Ramp Automated Engineering With AI Agents
7/10TL;DR
Ramp engineers report major productivity gains after deploying AI coding agents that autonomously optimize systems, reduce CI times, and operate across the full software lifecycle.
KEY POINTS
AI agents cut CI time by two-thirds
Ramp reduced continuous integration median runtime from 18 minutes to 6 minutes after deploying an autonomous optimization workflow. The system continuously profiled performance, applied iterative improvements, and validated results in production over several days without manual intervention.
Large-scale refactoring handled autonomously
AI models were tasked with complex codebase challenges, including eliminating import cycles and introducing lazy loading across a large Python monolith. Significant portions of these changes were merged, demonstrating the models’ ability to handle deeply structural engineering problems.
Shift from manual workflows to autonomous systems
Engineers increasingly rely on “dynamic workflows,” where AI agents determine their own steps rather than following predefined sequences. These systems dynamically allocate compute and reasoning, enabling them to tackle unpredictable tasks like system optimization.
Agents operate across the entire engineering lifecycle
AI tools are embedded from idea generation to production monitoring. They can write code, review pull requests, investigate incidents, respond to support tickets, and analyze logs using integrations with platforms like GitHub, Slack, Datadog, and Sentry.
Internal platforms enable widespread adoption
Systems such as Inspect function as “digital coworkers,” capable of fixing bugs, generating pull requests, and resolving incidents. Another interface, Glass, allows non-technical staff to interact with coding agents, expanding access beyond engineering teams.
Automation increasingly triggered without humans
More agent sessions are now initiated by automated triggers rather than direct user input. These include scheduled tasks, system alerts, and external events, signaling a shift toward continuous, background AI operations.
Security and safeguards built into infrastructure
The company applies principle of least privilege, giving agents restricted, read-only access where appropriate. Multiple layers of safeguards and close collaboration with security teams enable broader AI autonomy while limiting risk.
Focus on trace-based evaluation over benchmarks
Instead of relying solely on aggregate benchmarks, engineers analyze individual execution traces to determine whether agents follow the correct steps. This approach helps identify missing context, tools, or permissions.
Cultural shift favors experimentation and minimal constraints
Engineers are not restricted by token or cost budgets, encouraging exploration of high-impact use cases. Spending is evaluated based on return on investment, with the expectation that increased intelligence will drive higher productivity.
Designing for future models rather than current limits
Teams intentionally build systems anticipating rapid improvements in AI capabilities. This reduces the risk of short-lived infrastructure and allows faster adoption of new model releases.
CONCLUSION
Ramp’s experience highlights a transition toward autonomous AI agents acting as full participants in software development, with measurable efficiency gains and a growing emphasis on designing systems for rapidly advancing model capabilities.
Explain this