StrongDM Runs a Dark Factory With 3 Engineers. We Run a Dark Fund With 2 People.

Last week, Andrej Karpathy gave a talk at Sequoia Ascent 2026 that crystallised something I’ve been experiencing since February.

He retired the term “vibe coding.” The new frame: agentic engineering. Not amateurs prompting ChatGPT to spit out code (that’s vibe coding — it raises the floor). Agentic engineering is the professional discipline of coordinating fallible AI agents while preserving correctness, security, taste, and judgment.

Then he said the line that stopped me: “You can outsource your thinking, but you can’t outsource your understanding.”

I want to unpack that — not through the lens of software engineering, but through something most people haven’t considered: what happens when you apply these same principles to the entire knowledge work stack?

StrongDM’s Dark Factory

Dan Shapiro (CEO of Glowforge, Wharton Research Fellow) coined the term “dark factory” for software — borrowing from manufacturing, where robots work in unlit facilities because robots don’t need to see. He profiled StrongDM as the exemplar. Justin McCarthy and his three-person team at StrongDM ship production software — thousands of lines of Rust and Go — without anyone on the team writing a line of it. No one reviews it either. Two rules: “Code must not be written by humans” and “Code must not be reviewed by humans.” The workflow:

Specifications written in markdown → agents build → “digital twin universes” (behavioural clones of Okta, Jira, Slack) validate through scenario testing → LLM evaluators assess outcomes → humans approve the final result.

Shapiro mapped five levels of AI coding. Level 0 is no AI. Level 1 is autocomplete. Level 2 is where 90% of developers are stuck: AI writes code, human reviews every diff. Level 3: AI writes and tests, human reviews outcomes. Level 4: AI handles full features with scenario validation. Level 5: the dark factory.

The key insight most people missed: the bottleneck didn’t just shift from “writing code” to “describing what to build.” It shifted to something much bigger.

The Bigger Canvas

Here’s where I diverge from the pure engineering narrative.

If you apply first principles — not just to software, but to any complex knowledge work — the shift is:

Think dramatically bigger outcomes. When the marginal cost of analysis approaches zero, you can create products and services that would never have justified the investment before. We built a knowledge graph of 87,000+ entities — companies, investors, relationships, postmortems, analyses — that grows by hundreds daily. No fund our size would have attempted this pre-2025. The economics didn’t work. Now they do.

Reimagine entire workflows from the ground up. Not “make the existing process faster.” Throw out the existing process. Traditional deal screening: partner gets an intro, assigns an analyst, analyst spends a week on a memo, partner reads it. Our approach: the knowledge graph already has the company, its competitive landscape, funding history, and 64,800+ relationship connections before anyone writes a word. The Analytical Council runs multiple independent lenses simultaneously in structured batches, prioritised by deal relevance. A multi-dimensional assessment that would take a team a week happens in hours.

Design commercial models around the new math. This is the part nobody talks about. If a two-person operation can deliver analytical depth that previously required a team of twenty, the question isn’t “charge less per analysis.” The question is: how do you capture a share of the dramatically larger value created?

The Manthan-Tavaga relationship is proof of concept. Manthan Intelligence (the tech platform) enables Tavaga Fund to operate with analytical capacity that rivals firms with ten times the headcount. Tavaga introduces Manthan to portfolio companies for AI-native transformation. Manthan participates in the equity upside. The old model: charge £500/hour for analysis. The new model: share in the outcomes your intelligence infrastructure makes possible.

What the “Dark Fund” Actually Looks Like

Here’s the reality of running a two-person AI-native investment operation since February 2026.

The intelligence layer runs 24/7. A 40-page morning brief synthesised from 36+ sources arrives before I wake up. Daily deal scrapers pull funding announcements. Company enrichment agents fill in gaps. Postmortem scrapers document failures. The knowledge graph is updated continuously — it’s currently at 87,016 entities across 14 categories.

The analytical layer is calibrated, not just fast. Over 200 graded scorecards — real predictions about real companies, scored against actual outcomes. Weighted accuracy: 67.3%, improving from 60.9% three weeks ago. The system tracks every mistake, files a learning entry, and adjusts. The accuracy on decline detection is 100% — when the system identifies a company heading for trouble, it has been right every time in the scored dataset.

The content layer is autonomous. Charaka Notes on getmanthan.com publishes data-driven research five days a week. The system identifies patterns from the knowledge graph, drafts analysis, fact-checks against sources, and publishes. Twenty-plus pieces live. An autonomous engine — not a human typing articles.

The deal flow layer is intelligent. Not waiting for warm intros. Actively scanning, filtering, prioritising — based on pattern recognition across 13,800+ companies and 303 documented deaths.

Karpathy’s Framework Applied Beyond Software

Karpathy proposed that AI automates what you can verify. Traditional software automated what you could specify. The new paradigm automates what you can build a feedback loop around.

Investment analysis is deeply verifiable. You make a prediction (invest/pass), time passes, outcomes are observable (company raised, company died, company stagnated). That’s a verification loop. Which means you can calibrate. Which means accuracy improves systematically.

We’re running exactly that loop. Every scorecard is a training signal. The 67.3% accuracy isn’t the end state — it’s the current position on a curve that’s been improving 3+ percentage points per week.

The same principle applies to every knowledge work domain with observable outcomes: consulting recommendations, legal risk assessments, medical diagnoses, audit findings, strategic plans. If you can score the prediction against reality, you can build an agentic engineering system around it.

What I’m Building Toward

The thesis is simple: the future of knowledge work is being redesigned globally. AI-native architecture lets small teams deliver output at a scale that was previously impossible. India’s most AI-native investment operation is one proof point.

But this isn’t just a fund story. The architecture — knowledge graphs, multi-dimensional analytical councils, autonomous content engines, calibration loops — is applicable to any professional service where human judgment is the bottleneck and outcomes are measurable.

Karpathy’s closing line resonates: the work itself is being reorganised around agents. Define the context. Define the tools. Define the feedback loop. Define the guardrails. Let agents work. Preserve human understanding.

That’s what we’re doing. I’ll be documenting the journey here — what works, what breaks, what the numbers actually say.

Mayank Mathur | Founder, Manthan Intelligence | GP, Tavaga Fund

Read the full analysis on Charaka Notes.