Show Notes
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.
We also discuss: Under the hood of Ramp Sheets Inspect, Ramp's internal coding agent, and the self-improving monitor loop it powers Why finance professionals rejected code gen as too "black box" Why Anthropic models tend to excel at agentic spreadsheet manipulation The case for putting the agent outside the sandbox, not inside it The Loom-to-Markov-diagram process mining pipeline RLMs and how subagents can share memory in latent space Latent Briefing and KV-cache communication between subagents Reviving Golden Gate Claude with steering vectors on Gemma
Referenced: Alex Levinson Anthropic Ben Geist Claude Efficient Memory Sharing for Multi-Agent Systems via KV Cache Compaction (Ben Geist) Gemma Golden Gate Claude Graphviz Inspect Latent Briefing Loom Modal OpenAI Opus Qwen Ramp Ramp Labs Ramp Sheets Recursive Language Models (Alex Zhang) Retool Self-maintaining Ramp Sheets Steer AI
Where to find Alex: LinkedIn Twitter/X Website
Where to find Harrison: LinkedIn Twitter/X
Where to find LangChain: Website Docs
Send feedback or questions to [email protected]
Timestamps: (00:00) Introduction (01:13) The origin of Ramp Sheets (02:27) The Loom-to-Markov-diagram process mining pipeline (04:28) Why code gen approaches felt too "black box" to finance (06:13) Meeting finance where they already are: inside the spreadsheet (09:08) How far process mining got them (10:31 )Text descriptions and Graphviz DAGs as output (12:41) Under the hood of Ramp Sheets (14:52) Why the agent uses Python only as an escape hatch (15:47) Why Anthropic models excel at agentic spreadsheet manipulation (17:12) Frankensteining the OpenAI Agents SDK (17:43) The Ramp Sheets UX and fast vs. expert mode (19:58) Agent in a sandbox vs. agent with a sandbox (21:55) Vibe evals with expert humans (23:40) Inspect, the internal coding agent (24:13) The self-monitoring loop and auto-PRs (28:01) Other wacky experiments on Sheets (28:43) Memory experiments that didn't pan out (31:16) Latent Briefing and KV-cache subagent communication (35:13) Reviving Golden Gate Claude (37:47) Contrastive pairs and steering vectors (39:47) Picking the right layers in Gemma (41:37) What Ramp Labs looks for when hiring
More from Max Agency
Connor Heggie spent his early career on a fifteen-person self-driving startup run like a research lab, then moved to Scale AI's mapping team, before becoming the co-founder and CTO at Unify. Unify builds agents for go-to-market teams, and for the last two years, one of AI's big mainstream narratives has been automating away the sales rep entirely. Connor and his team built the opposite: an agent that gives every sales rep "an engineer in their back pocket." He walks through how Unify's harness evolved from million-agent batch jobs to a chat product, and unpacks the engineering that makes it cost-effective to run at scale. – We also discuss: How Unify cut 90-95% of costs two weeks before launch The 15-requests-per-second ceiling inside OpenAI's prompt cache Why your LLM judge must be a different model family The Speed Audit: why one at a time beats a table of 1,000 Why Unify's subagents are just a function call What working on self-driving taught Connor about running evals – Timestamps: 00:00 Introduction 01:30 "Go-to-market is a search problem" 06:10 The old workflow: drag-and-drop nodes over a million-row table 09:40 Why the harness is similar to a coding agent's 10:50 Running durable agents in the cloud without a full VM 13:15 Giving models pandas-like superpowers over a live table 15:50 Why Unify's subagents are just a function call 18:25 The 15-requests-per-second limit hiding inside OpenAI's cache 24:20 Optimizing for prompt caching hit rates 28:20 Fork versus child subagents 32:35 The Speed Audit: why one at a time beats a table of 1,000 37:50 Locking memory to keys instead of letting the agent freestyle 44:15 Why AI isn’t taking over sales 49:00 What working on self-driving taught Connor about running evals 52:46 Why your LLM judge must be a different model family 54:45 Ditching full VMs for Monty, a Python REPL that suspends 58:55 Semantic merge sort: why Connor is obsessed with RLMs 59:53 How Unify cut 90-95% of costs two weeks before launch 1:01:15 The case for "semantic linters" over skill files 1:04:55 Why Unify runs mostly on OpenAI, their first investor 1:06:12 Why 10x cheaper tokens still lose on tool efficiency 1:07:24 Why open-source models don’t make economic sense (yet) – Referenced: Anthropic ChatGPT Claude Claude Fable 5 Claude Sonnet 5 Context Hub Deep Agents GLM-5.2 GPT-5.4 Helm.ai LangSmith OpenAI QuickJS Recursive Language Models (RLM) Salesforce Scale AI Scaling Managed Agents Unify – Where to find Connor: Twitter/X LinkedIn – Where to find Harrison: Twitter/X LinkedIn – Where to find LangChain: Twitter/X LinkedIn – Send feedback or questions to [email protected]
Russell Kaplan is the President of Cognition, the $26 billion company behind Devin, the AI software engineer. He started his career in machine learning at Tesla's Autopilot team, then led the ML org at Scale AI, which had acquired his computer vision startup Helia. In this conversation, Russell unpacks why more and more engineering work is becoming "intelligence saturated," how the sidekick architecture inside Devin Fusion cuts cost without cutting quality, and why Cognition is underwriting a $10 million productivity guarantee. – We also discuss: Why mergeability is the next eval bar When to use Fable vs GPT models How smarter routing unlocked 35% better price performance Why letting users pick their own model is a UX bug Wiring agents to be proactive instead of reactive How Cognition scores a session productive or unproductive – Timestamps: (00:00) Introduction (01:21) Launching Devin at 13% on SWE-bench (02:47) When Devin became the number one committer to Devin (05:56) Intelligence saturation: when speed and cost start to matter more (07:05) When to use Fable vs GPT models (08:41) Why mergeability is the next eval bar (10:51) How open source maintainers shaped FrontierCode (14:56) The telltale sign of agent-written code (17:56) When per-person token spend starts to eclipse salaries (20:26) Why letting users pick their own model is a UX bug (21:46) How smarter routing unlocked 35% better price performance (23:50) Why Cognition still trains its own models (27:13) Deploying Cerebras at scale at 950 tokens per second (28:21) The Tesla chip debate: does an inference chip need division? (30:06) Wiring agents to be proactive instead of reactive (32:19) Why "software factory" is the wrong term (35:16) Why everyone is now a new grad (41:10) Forward deployed engineering, and why ex-founders thrive (43:16) The industry's incentive to get you tokenmaxxing (45:11) How Cognition scores a session productive or unproductive (46:51) Underwriting a $10 million productivity guarantee (48:11) Using productivity dashboards to coach, not police – References: Anthropic Cerebras Claude Fable Claude Opus Cognition Cognition's AI Productivity Guarantee Devin Devin Desktop Devin Fusion Devin Review Devin Security Swarm Estimating the Productivity of an Autonomous AI Software Engineer FrontierCode GPT-5.5 GPT-5.5-Cyber Harbor LangSmith OpenAI OpenRouter Model Fusion SWE-1.6 SWE-bench Tesla – Where to find Russell: LinkedIn Twitter/X – Where to find Harrison: LinkedIn Twitter/X – Where to find LangChain: Website Docs – Send feedback or questions to [email protected]
Eno Reyes is the co-founder and CTO at Factory, a $1.5 billion company turning signal into deployed code inside some of the world's biggest engineering orgs. Before founding Factory in 2023, he was an engineer at Microsoft and then Hugging Face. In this conversation, Eno unpacks why the harness matters more than the model running underneath it, how to build your own 24/7 software factory, and why he's "bullish on humans in the loop for a very long time.” – We also discuss: Why product management isn't going away The engineer who lint-checks Factory's own agents Why coding agents might become the best general agents The platonic representation hypothesis Why Factory might hide when "memory" is happening Building Factory's universal meta harness Tokenomics, and how routing cuts the bill – Timestamps: (00:00) Introduction (01:19) What a 24/7 autonomous software factory actually means (03:16) Why product management isn't going away (06:56) Alvin, the engineer who lint-checks Factory's own agents (10:16) "It makes me very bullish on humans in the loop for a very long time" (11:37) The Disney Epcot analogy for rolling out AI (16:47) Why coding agents might become the best general agents (20:54) The case against a model-independent harness, and Eno's counter (25:44) The platonic representation hypothesis explained (32:42) Why model quirks are like being left or right-handed (39:28) Why you could technically decompile Factory's entire harness (44:20) Why memory might be AI's most overused word (46:47) Inside AutoWiki and its Lore feature (55:32) Agent readiness: the deterministic feedback agents need (57:12) Missions: Factory's universal meta harness (1:01:12) "It's kind of turtles all the way down": validating the validators (1:04:37) Tokenomics: what missions cost, and how routing cuts the bill (1:11:08) Why Eno is bullish on open models (1:14:30) BenchBench, and why code review benchmarks might be broken – References: Aider Alvin Sng Amp Andrej Karpathy Anthropic Anthropic: Emotion concepts and their function in a large language model Cursor Deep Agents DeepWiki Epcot Factory GLM Harbor Hugging Face HuggingGPT Kimi LangGraph LangSmith MiniMax Open Knowledge Format (OKF) OpenAI OpenRouter Model Fusion Ramp Sakana Fugu SWE-bench Terminal-Bench – Where to find Eno: LinkedIn Twitter/X – Where to find Harrison: LinkedIn Twitter/X – Where to find LangChain: Website Docs – Send feedback or questions to [email protected]
Ben Tannyhill is a product manager at LangChain, where he's building LangSmith Engine—an agent that finds and fixes your agent's failures. Engine continuously analyzes your production traces, clusters them into actionable issues, and opens pull requests to fix them. Engine's architecture is a lot like an org chart: a main model delegating to a team of cheaper, faster sub-agents. It launched in public beta at Interrupt 2026, and in this conversation, Ben unpacks why it uses a sandbox as a tool, how the team turned it into a self-improving agent that learns from its own traces, and the hard problem of testing a fix before it ships. – We also discuss: Why Engine is "the agent for agent engineers" Making LangSmith agent-native with condensed trace views Why the team keeps handing more control to the agent Inside Engine's four sub-agents: the screener, verifier, and more Giving Engine memory with an agent overview document How to keep an always-on agent from blowing the inference budget Where Insights, Polly, and Engine are converging – Timestamps: (00:00) Introduction (01:25) LangSmith 101 (02:22) Why Engine is "the agent for agent engineers" (03:49) Under the hood: Engine is a deep agent (06:08) Clustering millions of traces with condensed views (10:10) Why the team keeps handing more control to the agent (13:21) Why Engine uses a sandbox as a tool (14:11) Engine's four sub-agents and the org-chart analogy (16:51) Evals for Engine: IssueBench, Harbor, and synthetic environments (23:05) How Engine evolved: from noisy PRs to an issue inbox (25:56) Inside Engine's memory: the agent overview document (29:25) How to keep an always-on agent from blowing the inference budget (30:52) What models Engine uses (31:30) How Engine was rolled out: from Forge to public beta at Interrupt (34:18) Inside the two teams building Engine (35:53) Where Insights, Polly, and Engine are converging (40:06) The missing piece: testing a fix before it ships (42:22) Running a branched agent, and the write-access eval problem (46:35) Using Engine as long-term memory (47:39) Pointing Engine at coding-agent traces (48:49) Running Engine on Engine: the meta self-improvement loop – References: Anthropic Chat LangChain Claude Code Claude Haiku Claude Opus Codex Context Hub Credit Genie Deep Agents Gemini GPT-5.5 Harbor Hex Insights Interrupt LangGraph LangSmith LangSmith Chat (formerly Polly) LangSmith Engine LangSmith Observability Mintlify OpenAI Palash Shah Terminal-Bench Unify – Where to find Ben: LinkedIn Twitter/X – Where to find Harrison: LinkedIn Twitter/X – Where to find LangChain: Website Docs – Send feedback or questions to [email protected]
Zack Reneau-Wedeen is the Head of Product at Sierra, the conversational AI platform behind customer-facing agents for most of the Fortune 20. Before Sierra, he spent seven years at Google as the founding PM for Google Lens and Google Podcasts, then led product at Robinhood and CoinTracker. Sierra is mostly known for customer support, but Zack reveals how and why the company is building agents that span the entire customer lifecycle, from browsing and booking to sales and loyalty. In this conversation, he argues agentic commerce will be bigger than e-commerce, explains why he's a "monolith loyalist", and unpacks why, when a model looks dumb, the problem is usually you. – We also discuss: How Sierra's no-code layer compiles down to agent code, and back again Why most multi-agent systems just ship your org chart Inside Sierra's modular voice architecture: thinking, listening, and talking in parallel Why Sierra built a PCI-certified stack for voice payments How outcome-based pricing aligns incentives Why there's no breakout memory company – Timestamps: (00:00) Introduction (03:39) Analyze, build, release: how you build on Sierra (07:54) Inside Ghostwriter (11:04) Meeting models on their turf “80% of the time (17:47) The one constraint Claude Code doesn't have (19:35) Agent-to-agent: when an API call still beats MCP (21:02) Why agentic commerce will be bigger than e-commerce (27:31) Running models in parallel and ensembling transcription (32:22) Inside the Agent Data Platform (40:00) Context engineering: everything it needs, nothing more (41:38) "Whenever you think the model's too dumb, the model's actually too smart" (46:13) Why multi-agent systems are a trap (48:44) Voice 101: latency, naturalism, and 60 languages (56:11) When voice-to-voice passes 50%: the over/under (57:03) Making memory a first-class primitive (1:02:47) Why there's no breakout memory company (1:08:02) Why the solution to all AI problems "is more AI" (1:09:20) Why Sierra open-sources the tau-bench universe (1:14:42) How outcome-based pricing aligns incentives (1:20:26) Who thrives as a forward-deployed agent builder (1:22:16) The Formula One analogy: why product is the bottleneck (1:25:47) How Sierra interviews for agency – References: Agent2Agent (A2A) Protocol Anthropic ChatGPT Claude Claude Code Claude Mythos Claude Opus 4.5 Codex Deep Agents Gemini Hawaiian Airlines LangGraph Model Context Protocol (MCP) Not Another Workflow Builder Redfin Sentry Shopify Silero SiriusXM Stripe Tau-bench Thinking Machines Lab – Where to find Zack: LinkedIn Twitter/X Sierra – Where to find Harrison: LinkedIn Twitter/X – Where to find LangChain: Website Docs – Send feedback or questions to [email protected]