A talk for future venture capitalists

Becoming an
Agentic VC

How to be the most valuable non-engineer at a venture fund.

Avi Flombaum · Innovent Capital · avi.nyc

Who I am

An unlikely path into venture.

  1. College dropout who taught himself to code
  2. Startups, then founded Flatiron School
  3. WeWork
  4. Now: venture capital

Today at Innovent Capital — a 6-person family office, ~26 portfolio companies over 3 years. I'm the only programmer on the team. That's the point.

The provocation

The old path into VC was a finance degree and a banking job. The new edge is being the person who can make the whole firm 10× more capable with AI.

A quick survey

What do you actually use?

Hands up — honestly.

  • ChatGPT
  • Claude
  • Claude Code / Claude Co-Work
  • Specialty models (ElevenLabs, image & video…)

The gap

Most people use 5% of the tool.

A chatbot is: ask a question, get an answer, copy–paste. Useful — but it's a window you type into.

This talk is the other 95%: AI as a system that does work.

Know the tool · 1

What AI is good at.

  • Defining and following a process
  • Searching & finding information
  • Reading messy inputs — PDFs, emails, transcripts
  • Correlating, normalizing, summarizing
  • Drafting and reformatting into new outputs

…and doing all of it tirelessly, in parallel, on a schedule.

Know the tool · 2

What AI is bad at.

  • Knowing your firm's private context, unless you give it
  • Perfect accuracy on high-stakes numbers
  • Knowing when it's wrong (it makes things up, confidently)
  • Long-horizon judgment and taste

This is why harnesses exist — structure around the model covers its weak spots.

Vocabulary

Prompt

The instruction you give the model.

A good one = role + task + context + output format.

Most people stop here. A prompt is just one move.

Vocabulary

Context

Everything the model can “see” while it works.

Your files, your data, prior messages, your instructions.

The firm's proprietary data is the moat — the thing ChatGPT-on-the-internet can't touch.

Vocabulary

Loop

The agent acts, sees the result, and acts again — until the task is done.

Read → extract → check → store → repeat for 50 files.

This is what turns “chat” into work.

Vocabulary

Skill & memory.

A skill is a reusable instruction sheet you write once — so the AI never re-figures-it-out.

“When you get a company update, always look for MRR, burn, runway, headcount, and asks.”

Memory = the data & decisions that persist between runs. Teach once, reuse forever.

The big word

An agentic harness.

An application around the model that can: take input → run workflows → use skills & memory → produce artifacts → improve itself.

Claude Code and Claude Co-Work are harnesses. So is anything you build.

Not the only option — open models, OpenRouter, dropping costs. Don't memorize the tool. Understand the shape.

The role

The job is consulting, not coding.

Walk the firm. Ask the questions an outside consultant would:

  • What's painful, repeated, manual?
  • What's time-consuming every week?
  • What proprietary data do we have that ChatGPT doesn't?

You don't write code. You see the workflow and design the harness.

The mindset

AI is a force multiplier, not a replacement. You build the harness and hand it to the people who do the work. Your value is leverage for the whole team.

The mental move

“What would a human do?”

Break the task into human steps → map each to what AI is good at → wrap it in one frame:

  1. 01Workflowwhat a human does
  2. 02Datathe firm's context
  3. 03Insightsurfaced & acted on

The method

Start with a workbench.

One folder that holds all the context, all the prompts, all the instructions for a task.

Everything the AI needs, in one place. That's a project.

A worked example

26 companies. Every month, founders send updates — a PDF, an email, a spreadsheet, sometimes a phone call. The team compiles it all and reports up to the LP.

How do you turn that pile into a system?

Step 1 · Workflow

Input + a parsing skill.

Where's the input? A watched folder, a Dropbox, a Slack channel people post to.

The skill: every update has MRR, burn, runway, headcount, wins, risks, asks. Define it once — the AI never has to guess.

Step 2 · Data

Store & normalize.

  • Raw input → structured output (keep both, like Granola does for meetings)
  • Stored over time, correlated to the right company — a time series
  • Markdown files? A Google Sheet? A database?

A valid move: ask Claude which is best. You don't have to know everything.

Step 3 · Insight

Surface it & act.

Split the quantitative (metrics) from the qualitative (asks):

“We're hiring a CMO.”   “Know any enterprise sales leads?”   “We're raising.”

A harness doesn't just store — it alerts and acts.

The building blocks

Six primitives. That's the kit.

  • 01Loops
  • 02Skills
  • 03Agents
  • 04Scheduled tasks
  • 05Memory
  • 06Artifacts

You don't rebuild from scratch. You teach the AI once.

Your move

What does the market-research button look like? The competitor-research button? The due-diligence button?

Every firm is unique. The proprietary data is the moat.

How to start

Start small. Ship one button.

  1. Pick one painful, repeated workflow
  2. Ask Claude how to structure it
  3. Build the smallest harness that works
  4. Capture it as a skill — so you never redo it
  5. Iterate. Add the next button.

The career angle

You're not competing with the analysts. You're multiplying them. Be the person who builds the firm's leverage.

Recap

The whole talk in one frame.

  1. 01Workflow
  2. 02Data
  3. 03Insight
  • Loops
  • Skills
  • Agents
  • Scheduled tasks
  • Memory
  • Artifacts

Play with Claude Code / Co-Work. Build one small thing this week.

Thank you

What button will
you build?

avi.nyc · im@avi.nyc · @aviflombaum

Let's talk.  Questions?

Appendix · what we run at Innovent

Three harnesses. One framework.

PostMoney

The update harness, productionized.

VCMatcher

Research & matching at machine speed.

EOW

The report that writes itself.

Each one = workflow → data → insight.

Appendix · 01

PostMoney

Founder updates → portfolio intelligence

Workflow
Drop in a PDF / email — it resolves the company & period.
Data
Extracts ARR, burn, headcount with confidence + provenance.
Insight
“Show every company with <6 months runway.”

Parallel AI jobs · prompt caching for cost · a human signs off.

Appendix · 02

VCMatcher

Research & matching at machine speed

Workflow
Scrape a VC site + parse a pitch deck → structured profiles of both.
Data
Embeddings + sector weighting → a tunable match score.
Insight
A match brief: who to approach, what to say, what to fix first.

3–4 weeks of analyst research → hours. Runs on the firm's data.

Appendix · 03

EOW

The end-of-week report that writes itself

Workflow
Friday: agents fan out across git, Asana & files in parallel.
Data
One shared registry tells every agent what matters. Teach once.
Insight
A dated exec report, synced to Notion for leadership.

Loops + skills + agents + schedule + memory + artifact — all in one.

Appendix · what they share

None of them is a chatbot.

  • A harness around real, proprietary firm data
  • Replaces hours of manual work
  • Produces an artifact a non-technical partner uses

All from the same three questions: what's the workflow, where's the data, where's the insight?

Becoming an Agentic VC Innovent Capital