Sign up to save your podcasts
Or
Share How a VC runs his investment workflow on AI agents
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
How a VC runs his investment workflow on AI agents
Discover how a seed-stage investor turned a slow, manual research process into an AI analyst he runs from a single Slack message. This deep dive walks through why most funds won’t let AI near their confidential deal documents and why he did it anyway, the exact architecture that stops the agent from inventing numbers on financial data, and a live demo of the whole thing pulling a full company brief from four connected systems in seconds.
In this episode:
* Why trusting AI with confidential data is really a question about the vendor, not the technology
* How he built the first working version over a weekend, and why it is never actually finished
* The four narrow sub-agents behind the analyst, and why narrow scope beats one agent doing everything
* The rule that keeps it honest: clickable sources, a human review step, and treating the agent like an intern
* When to call a plain function instead of letting the model use judgment, so it stops making things up
* Why the near future is agent to agent, not human to human, and what that changes for founders
* How to build your own, whether you want a frontier model’s harness or full open-source control
Timestamps:
(00:00) Introduction
(00:40) The problem: Portfolio answers scattered across systems
(02:49) Why let AI near confidential deal documents at all
(04:11) Do you actually trust OpenAI and Anthropic with your data?
(04:26) Choosing a harness and locking it down with Tailscale
(07:36) Building the first version over a weekend
(08:47) Live demo: a full company brief from one Slack DM
(10:48) Why the near future is agent to agent
(12:30) Verification: clickable sources and the intern rule
(15:14) The architecture: Claude Code and the Agent SDK
(18:24) Compounding org knowledge with ByteRover
(21:18) The four sub agents and the data cleanup grind
(24:43) When to call a function instead of trusting LLM judgment
(28:44) Loops, goals, and orchestrator burnout
(29:47) Advice for building your own AI analyst
(32:51) Turn paranoia into verification, then just build
(34:06) Wrap up
Resources & Links:
* Claude
* Attio’s docs built for AI agents: https://attio.com/llms.txt
Connect with Binh Tran:
* avv.co
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com
View all episodes
By Shipping with AIDiscover how a seed-stage investor turned a slow, manual research process into an AI analyst he runs from a single Slack message. This deep dive walks through why most funds won’t let AI near their confidential deal documents and why he did it anyway, the exact architecture that stops the agent from inventing numbers on financial data, and a live demo of the whole thing pulling a full company brief from four connected systems in seconds.
In this episode:
* Why trusting AI with confidential data is really a question about the vendor, not the technology
* How he built the first working version over a weekend, and why it is never actually finished
* The four narrow sub-agents behind the analyst, and why narrow scope beats one agent doing everything
* The rule that keeps it honest: clickable sources, a human review step, and treating the agent like an intern
* When to call a plain function instead of letting the model use judgment, so it stops making things up
* Why the near future is agent to agent, not human to human, and what that changes for founders
* How to build your own, whether you want a frontier model’s harness or full open-source control
Timestamps:
(00:00) Introduction
(00:40) The problem: Portfolio answers scattered across systems
(02:49) Why let AI near confidential deal documents at all
(04:11) Do you actually trust OpenAI and Anthropic with your data?
(04:26) Choosing a harness and locking it down with Tailscale
(07:36) Building the first version over a weekend
(08:47) Live demo: a full company brief from one Slack DM
(10:48) Why the near future is agent to agent
(12:30) Verification: clickable sources and the intern rule
(15:14) The architecture: Claude Code and the Agent SDK
(18:24) Compounding org knowledge with ByteRover
(21:18) The four sub agents and the data cleanup grind
(24:43) When to call a function instead of trusting LLM judgment
(28:44) Loops, goals, and orchestrator burnout
(29:47) Advice for building your own AI analyst
(32:51) Turn paranoia into verification, then just build
(34:06) Wrap up
Resources & Links:
* Claude
* Attio’s docs built for AI agents: https://attio.com/llms.txt
Connect with Binh Tran:
* avv.co
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit shippingwithai.substack.com