Avsnitt
-
Nick Larus-Stone is the Head of AI at Benchling, the R&D data platform that life science companies use to store and manage their experiments, samples, instruments, and analysis. Benchling has been around for since 2012. In October 2025, it launched Benchling AI, an intelligence layer with a chat interface, backed by an agent, that helps scientists find data, design experiments, and write reports. Nick came to Benchling through its acquisition of Sphinx Bio, the analysis startup he founded. In this conversation, Nick walks through what it takes to build agents for scientific work, and where the playbook from coding agents holds up and where it breaks down.
–
We also discuss:
Why Benchling invests so heavily in getting clean data upfrontHow they cross-check answers between models to get more out of each oneWhy and how Benchling leans on production tracesWhere AI actually helps science today, and where it still gets stuckWhy understanding LLMs is closer to biology than software engineering–
Timestamps:
00:00 Intro
01:22 What Benchling AI is, and the 14-year data platform underneath it
04:36 Why a decade of structured data is a core advantage
05:57 The architecture under the hood
08:28 Similarities and differences compared to a coding harness
11:14 Benchling’s multi-agent architectures
14:36 Dealing with verifiable vs non-verifiable tasks
16:19 Doing evals when clean benchmarks aren’t possible
18:13 Context engineering: SQL vs. file-based harnesses
22:11 Memory: agents that create and update their own skills
25:30 What user education for scientists looks like
30:33 Why understanding LLMs is closer to biology than software
33:28 When will agents discover a novel cure for disease?
44:58 The future of harnesses in science
48:13 Why fine-tuning on biology hasn't beaten frontier models
–
References:
Agent Skills (Claude Docs)Benchling’s Deep Research AgentClaude (Anthropic)Design of experiments (DOE)FDA Investigational New Drug (IND) applicationGemini (Google)Google AI co-scientistLangSmithModel Context Protocol (MCP)The Ralph (Wiggum) Loop (Geoffrey Huntley)Sphinx Bio–
Where to find Nick:
BenchlingLinkedInTwitter/X–
Where to find Harrison:
LinkedInTwitter/X–
Where to find LangChain:
WebsiteDocs–
Send feedback or questions to [email protected]
-
Geng Sng is co-founder and CTO of Cogent, which builds autonomous agents that remediate vulnerabilities for enterprise security teams. Today, Cogent's agents process billions of security events per day, maintaining a live context graph of every asset and vulnerability across customer environments. In this conversation, Geng walks through Cogent's hot vs cold context split, the sub-agents that handle side quests, and the two graphs they run in parallel.
–
We also discuss:
Why defensive security is harder for AI than offensiveUnder the hood of Cogent's three agentsInside Cogent's “read only” by-default sandboxesWhy graph databases don't scale for security dataCogent Research and the move into formal verificationWhy interactive agents need a deeper planning phase to one-shot–
Referenced:
Abnormal AIAmazon S3AnthropicBashChatGPTClaude CodeClaude MythosCodeMenderCodexCogentCursorGoogle DeepMindGPT-5.5-CyberJupyterLettaMozillaOpenAIOpus 4.6Opus 4.7Vercel–
Where to find Geng:
LinkedIn–
Where to find Harrison:
LinkedInTwitter/X–
Where to find LangChain:
WebsiteDocs–
Send feedback or questions to [email protected]
–
Timestamps:
00:00 Why mean time to exploit collapsed from years to minutes
02:08 Inside Cogent's Agent Lake architecture
05:11 Why Cogent rejected graph databases
10:48 The trust ladder before agents touch production
15:13 The three types of agents inside Cogent
17:07 How Cogent sandboxes its agents
19:16 Short-circuiting interactive agents with a deeper planning phase
24:31 What to do when users believe agents too much
31:21 Why sub-agents let agents go on side quests
34:59 Two-tiered evals and the metric that catches bad prompts
40:00 Cogent’s unique approach to context
48:39 Cogent Research and the move into formal verification
51:33 The single trait Cogent hires for
54:00 Open-sourcing models within six months
57:07 Why defensive security won’t be commoditized anytime soon
1:00:51 The founding insight behind Cogent
-
Saknas det avsnitt?
-
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 SheetsInspect, Ramp's internal coding agent, and the self-improving monitor loop it powersWhy finance professionals rejected code gen as too "black box"Why Anthropic models tend to excel at agentic spreadsheet manipulationThe case for putting the agent outside the sandbox, not inside itThe Loom-to-Markov-diagram process mining pipelineRLMs and how subagents can share memory in latent spaceLatent Briefing and KV-cache communication between subagentsReviving Golden Gate Claude with steering vectors on GemmaReferenced:
Alex LevinsonAnthropicBen GeistClaudeEfficient Memory Sharing for Multi-Agent Systems via KV Cache Compaction (Ben Geist)GemmaGolden Gate ClaudeGraphvizInspectLatent BriefingLoomModalOpenAIOpusQwenRampRamp LabsRamp SheetsRecursive Language Models (Alex Zhang)RetoolSelf-maintaining Ramp SheetsSteer AIWhere to find Alex:
LinkedInTwitter/XWebsiteWhere to find Harrison:
LinkedInTwitter/XWhere to find LangChain:
WebsiteDocsSend 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
-
Florian Juengermann is the co-founder and CTO of Listen, an AI startup that turns qualitative research across hundreds of interviews, surveys, and focus groups into structured, traceable insights. Listen's agents analyze responses at scale, and Florian has rearchitected the system multiple times to get there. In this conversation, he walks through the virtual table architecture at the core of their Research Agent, how small models run map-reduce classification across thousands of open-ended responses, and the self-reviewing feedback subagent that catches errors during long async runs.
We also discuss:
The three agents inside Listen's platformHow Listen rearchitected from a simple RAG bot to a multi-agent system multiple timesWhy the PowerPoint subagent was completely rebuilt using Claude's code SDKContextual prompt engineering as an alternative to skillsHow Listen keeps report numbers live as new interview responses come inWhen to trigger the long-running agent vs. showing early resultsWhat Florian looks for when hiring agent engineersReferences:
AnthropicChatGPTClaudeClaude Code SDKE2BEmotional IntelligenceGPT MiniHaikuListenOpenAIPandasPostgresPythonResearch AgentRenderZoomWhere to find Florian:
LinkedInTwitter/XWhere to find Harrison:
LinkedInTwitter/XWhere to find LangChain:
WebsiteDocsSend feedback or questions to [email protected]
Timestamps
00:00 Introduction
01:25 The three agents inside Listen's platform
03:15 Live chat vs. long async runs, and how Listen tunes for each
05:33 Under the hood of the Research Agent
06:37 Listen's virtual table architecture
07:34 How small models classify thousands of open-ended responses
10:05 Running code in a sandbox: how E2B fits in
11:52 Why Listen rebuilt the PowerPoint subagent from scratch
14:11 Contextual prompt engineering instead of skills
16:32 The feedback subagent that reviews its own reports
18:14 How Listen runs evals in production
19:47 Unexpected ways users push the agent to its limits
21:42 How many times Listen has rearchitected, and why
24:59 Trace observability: depth over breadth
26:10 Lessons from running Claude Code SDK inside E2B
27:42 Memory: what's solved and what isn't
29:10 The Composer agent UX: co-editing a document with AI
35:50 How Listen keeps report numbers live as new responses come in
43:47 What Listen looks for when hiring agent engineers
-
Izzy Miller is an AI engineer at Hex, an AI analytics platform that was one of the first companies to ship data agents to real paying users. Today, Hex runs a multi-agent system with nearly 100K tokens of tools, and Izzy is building a 90-day simulation to evaluate whether those agents actually get smarter over time. In this conversation, he walks through the harness decisions that shaped their architecture, the failure modes Hex is seeing at scale, and what it takes to build an eval that no current model can pass.
We also discuss:
Why data agents are harder to verify than coding agentsUnder the hood of Hex’s agentsHow Hex is unifying separate agentsWhy most eval sets are badThe 90-day simulation for long-horizon evalsHow Izzy went from marketing to AI engineerReferences:
Andon LabsAnthropicBarry McCardelChatGPTClaude CodeClaude Sonnet 4.6DBTGPT-3.5 TurboGPT-5.3 Codex SparkGPT-5.4HexLangChainLangSmithLookerOpenAIOpus 4.6Satya NadellaSnowflakeVending MachineWhere to find Izzy:
LinkedInTwitter/XWhere to find Harrison:
LinkedInTwitter/XWhere to find LangChain:
WebsiteDocsSend feedback or questions to [email protected]
Timestamps:
01:35 Where Hex's notebook agent started
03:46 The moment Hex knew it was time for agents
07:36 Why data agents are harder to verify than coding agents
09:30 How Hex is unifying separate agents
13:28 Under the hood of the notebook agent
15:41 The harness features that are now holding the agent back
17:41 Why Hex built their own orchestrator
18:59 Managing nearly 100K tokens of tools
20:49 Ephemeral queries and agent behavior trade-offs
24:46 The UX problem with showing agents' thinking
27:28 Why verification is harder than transparency for data agents
31:00 Memory, context conflicts, and collapse modes
34:38 How Hex built their internal eval system
39:29 Why most eval sets are bad
44:30 The 900% quota eval that every model fails
46:55 Model upgrades and the "in distribution" debate
51:34 How Izzy went from marketer to AI engineer
59:59 The 90-day simulation for long-horizon evals
-
Welcome to Max Agency, the podcast that goes deep into how the best agents are being built by builders like you. I'm Harrison Chase, CEO of LangChain, the agent engineering company, and I'll be your host.
