How Ramp built an AI agent that can think outside of tokens | Alex Shevchenko

<p>Alexander Shevchenko is the head of applied research at Ramp, where he leads Ramp Labs – the team behind Ramp Sheets and a steady stream of public AI engineering experiments. Ramp Sheets started as an internal process mining tool that turned Loom videos of accountants into Markov diagrams, before evolving into the agentic spreadsheet editor that shipped in November. In this conversation, Alex walks through the architecture under the hood, why Ramp biases the agent toward Excel formulas over Python code gen, and two recent Labs experiments: Latent Briefing and a user-steerable revival of Golden Gate Claude.</p><p><br></p><p><strong>We also discuss:</strong></p><ul><li>Under the hood of Ramp Sheets</li><li>Inspect, Ramp&#39;s internal coding agent, and the self-improving monitor loop it powers</li><li>Why finance professionals rejected code gen as too &quot;black box&quot;</li><li>Why Anthropic models tend to excel at agentic spreadsheet manipulation</li><li>The case for putting the agent outside the sandbox, not inside it</li><li>The Loom-to-Markov-diagram process mining pipeline</li><li>RLMs and how subagents can share memory in latent space</li><li>Latent Briefing and KV-cache communication between subagents</li><li>Reviving Golden Gate Claude with steering vectors on Gemma</li></ul><p><br></p><p><strong>Referenced:</strong></p><ul><li><a href="https://www.linkedin.com/in/alex-levinson/" target="_blank" rel="ugc noopener noreferrer">Alex Levinson</a></li><li><a href="https://www.anthropic.com/" target="_blank" rel="ugc noopener noreferrer">Anthropic</a></li><li><a href="https://www.linkedin.com/in/benjamin-geist/" target="_blank" rel="ugc noopener noreferrer">Ben Geist</a></li><li><a href="https://www.anthropic.com/claude" target="_blank" rel="ugc noopener noreferrer">Claude</a></li><li><a href="https://x.com/RampLabs/status/2042660310851449223" target="_blank" rel="ugc noopener noreferrer">Efficient Memory Sharing for Multi-Agent Systems via KV Cache Compaction (Ben Geist)</a></li><li><a href="https://ai.google.dev/gemma" target="_blank" rel="ugc noopener noreferrer">Gemma</a></li><li><a href="https://www.anthropic.com/news/golden-gate-claude" target="_blank" rel="ugc noopener noreferrer">Golden Gate Claude</a></li><li><a href="https://graphviz.org/" target="_blank" rel="ugc noopener noreferrer">Graphviz</a></li><li><a href="https://builders.ramp.com/post/why-we-built-our-background-agent" target="_blank" rel="ugc noopener noreferrer">Inspect</a></li><li><a href="https://x.com/RampLabs/status/2042672773747589588" target="_blank" rel="ugc noopener noreferrer">Latent Briefing</a></li><li><a href="https://www.loom.com/" target="_blank" rel="ugc noopener noreferrer">Loom</a></li><li><a href="https://modal.com/" target="_blank" rel="ugc noopener noreferrer">Modal</a></li><li><a href="https://openai.com/" target="_blank" rel="ugc noopener noreferrer">OpenAI</a></li><li><a href="https://www.anthropic.com/claude/opus" target="_blank" rel="ugc noopener noreferrer">Opus</a></li><li><a href="https://qwen.ai/" target="_blank" rel="ugc noopener noreferrer">Qwen</a></li><li><a href="https://ramp.com/" target="_blank" rel="ugc noopener noreferrer">Ramp</a></li><li><a href="https://ramplabs.substack.com/" target="_blank" rel="ugc noopener noreferrer">Ramp Labs</a></li><li><a href="https://labs.ramp.com/sheets" target="_blank" rel="ugc noopener noreferrer">Ramp Sheets</a></li><li><a href="https://alexzhang13.github.io/blog/2025/rlm/" target="_blank" rel="ugc noopener noreferrer">Recursive Language Models (Alex Zhang)</a></li><li><a href="https://retool.com/" target="_blank" rel="ugc noopener noreferrer">Retool</a></li><li><a href="https://ramplabs.substack.com/p/self-maintaining" target="_blank" rel="ugc noopener noreferrer">Self-maintaining Ramp Sheets</a></li><li><a href="https://labs.ramp.com/steer-ai" target="_blank" rel="ugc noopener noreferrer">Steer AI</a></li></ul><p><br></p><p><strong>Where to find Alex:</strong></p><ul><li><a hr