Ramp, the corporate spend management company, has deployed agentic AI to continuously monitor, triage, and propose fixes for its Ramp Sheets software, an AI-native spreadsheet editor.
The automated tool operates without human intervention, though human engineers review all code changes before deployment, the company announced on social media.
The engineering team built the agentic system using Ramp Inspect, an internal background coding application. Inspect creates a sandboxed development environment, allowing AI to execute API requests, run tests, and reproduce software defects against live code.
Developers initially programmed a scheduled nightly AI review to test core features and probe for hidden bugs. This approach successfully identified broad regressions but failed to isolate narrow, situational defects because the agent lacked a specific mission and followed repeating review paths.
“Prioritisation at production scale requires a level of intelligence surpassing that of any model available today,” the company stated regarding the limitations of AI. To resolve this, Ramp integrated its agents with Datadog, a cloud monitoring platform.
The revised system generates specific code monitors when engineers merge a pull request, which is a proposed code change. If a monitor detects abnormal metrics, a webhook activates an agent to reproduce the error and propose a solution.
The automated infrastructure expanded from 10 manually written monitors to over 1,000 AI-generated trackers, equating to one monitor for every 75 lines of code.
It identified 40 verified software bugs during its first week of operation. “Offloading this grunt work to an agent lets us ship faster,” the engineering team noted.
Engineers introduced an automated triage step to address false-positive alerts caused by poorly calibrated thresholds. When an alert triggers, the agent assesses the scope. It submits a fix and logs the code change link for genuine errors, while automatically tuning or deleting the monitor for irrelevant noise.
While OpenAI's GPT-5 series proved to be thorough debuggers, Ramp stated that Anthropic’s Opus 4.6 demonstrated greater accuracy as a triage evaluator and superior performance in filtering irrelevant alerts.
The company anticipates that AI models will eventually assume primary responsibility for software defence. Until reliability improves, Ramp maintains its traditional, manually written instrumentation alongside the automated systems to guard against severe incidents.
Recently, Ramp also expanded the capabilities of its Inspect agent.
The proprietary tool currently writes approximately 30% of all pull requests merged into the company's frontend and backend repositories. Engineers designed the system to support real-time, multiplayer collaboration across platforms like Slack, allowing multiple staff members to work within a single coding session simultaneously.
Ramp continues to utilise traditional, manually written instrumentation alongside the automated systems to guard against severe incidents.
“Auto-generated monitors are powerful but opaque, and not yet reliable enough to be your only line of defence,” the company explained, adding that this dynamic will shift as AI models improve.
Your reaction