💥Fund Performance Benchmarks, How Agent Loops Work, Portfolio Construction, Enterprise Software Playbook & More

---

1. Key Themes

Theme 1: The Power Law Extends Further Up the Valuation Stack Than Assumed

Late-stage and crossover investors have a data-backed argument for continued compounding at mega-scale. Coatue's Thomas Laffont presented data showing that the probability of reaching the next tier increases as companies get larger — not decreases.

"8% of unicorns ($1-10B) go on to reach decacorn status, 13% of decacorns ($10-100B) reach $100B, and 31% of $100B+ companies have gone on to 10x. The higher you are in the valuation stack, the more likely you are to keep compounding."

"A company already valued above $100B is nearly four times more likely to 10x than a unicorn is to reach decacorn status."

---

Theme 2: The Traditional Enterprise Software GTM Playbook Is Obsolete

The wedge-to-suite-to-platform model — the dominant framework for B2B software investing and company-building for two decades — has been invalidated by a new cohort of AI-native companies that collapsed the timeline from $0 to $100M ARR.

"Cursor, Clay, Harvey, Cognition, Sierra, Baseten, Fireworks, and Lovable all went from roughly $0 to $100M ARR in the past two years, compressing what used to be a 6-10 year journey."

"The traditional Act I harbor (a niche to reach $10-50M ARR before building adjacencies) now functions as a constraint. As software engineering costs drop, building Act I and Act II simultaneously is no longer unrealistic."

---

Theme 3: The 2021 Vintage Is a Cautionary Benchmark for Today's AI-Era Pricing

Carta's fund performance data provides the most rigorous empirical evidence yet that overpaying at entry — even in a bull market — destroys LP returns at scale. This is directly relevant to investors evaluating current AI entry prices.

"The 2021 vintage five-year distribution: 50th percentile 1.04x, 75th 1.29x, 90th 1.54x, 95th 1.93x… The entire cohort's distribution is compressed toward 1x in a way no prior vintage shows at the same age."

"The 2021 cohort's 95th percentile (1.93x) is below the 2017 cohort's 75th percentile (2.21x)."

---

Theme 4: AI Agent Infrastructure Is Nascent — Closed-Loop Architecture Is the Key Viability Test

The distinction between open and closed agent loops is an under-appreciated dividing line separating demo-worthy products from production-viable ones — and it has direct implications for how agent infrastructure companies should be evaluated.

"Open loops give the agent wide exploratory space. Holmberg notes they burn an 'insane amount of tokens' and without tight standards become a 'slop machine.' Currently viable only for teams with uncapped compute budgets."

"A closed loop: a human designs the path first: clear goal, defined steps, an eval gate at each step, a handoff point. The agent loops within that structure, and each run feeds performance data into the next. Runs on a normal budget and improves over time."

---

Theme 5: VC Portfolio Construction Math Has a Well-Defined Danger Zone

Monte Carlo simulation analysis from Gradient Ventures shows that extreme concentration — the instinct of many conviction-driven VCs — destroys median outcomes without proportionally improving upside, particularly at smaller fund sizes.

"Maximum concentration (22 companies, 15% ownership) has a median TVPI of 0.9x and a 90th percentile of 5.5x, below the less-concentrated lead strategy at the same percentile."

"At $30M, high-concentration strategies shrink the portfolio to 9-13 companies. Top quartile falls to 1.4-1.5x and top decile to 3.5x, below what more diversified strategies produce. Half of simulated maximum-concentration funds lose LP money."

---

2. Contrarian Perspectives

Perspective 1: Bigger VC Firms Are Scaling for Fundraising Convenience, Not Returns

The conventional wisdom is that larger, branded platform VC firms deliver superior returns through process, network, and resources. Dan Gray (Odin Times), citing Ewens and Rhodes-Kropf (SSRN), argues the opposite: firm scale actively dilutes the thing that generates returns.

"Value in VC firms resides in individual partners, not firm brand or process. Larger firms dilute individual partner influence and appear to scale primarily to reduce fundraising friction, not to improve returns."

"The VC industry has grown approximately 20x over two decades without a proportional increase in high-quality exits. Capital surges tend to intensify competition for the same technologies rather than fund genuinely new categories."

This is substantiated with an additional empirical point: "Excluding 2021, years with a larger share of capital going to first-time funds show a positive correlation with subsequent exit value, consistent with the thesis that emerging managers source outliers that incumbent firms miss."

---

Perspective 2: A Defensible Wedge Is Now a Liability, Not an Asset

The standard VC screening question — "what's your defensible wedge?" — may actively select against the companies most likely to win in the current AI environment. Conviction's Mike Vernal has explicitly updated his investment filter away from wedge-thinking.

"Vernal initially thought Cursor's plan to replace VS Code at seed stage was too aggressive. He was wrong, and now says replacing VS Code feels under-ambitious. Ideas that look too aggressive on day one may already be too conservative by fund close."

"His investment filter has shifted from 'what is the protective wedge' to 'unreasonable, unrelenting ambition,' because the compressed timeline rewards founders who plan to build the full stack from the start."

---

Perspective 3: Most "Autonomous Agent" Products Are Not Economically Viable for the Average Buyer

The prevailing narrative positions AI agents as transformative and near-term deployable across enterprises. Holmberg's framing suggests most current agent products are open-loop architectures that only make economic sense for a narrow set of well-capitalized buyers.

"Open loops… burn an 'insane amount of tokens' and without tight standards become a 'slop machine.' Currently viable only for teams with uncapped compute budgets."

The implication: the near-term infrastructure opportunity is in tools that make closed-loop construction faster, not in open-ended autonomous agents — a direct challenge to many current AI agent valuations.

---

3. Companies Identified

---

4. People Identified

---

5. Operating Insights

Insight 1: For Founders — Plan the Full Stack From Day One, Not the Wedge

The evidence from the AI-native $100M ARR cohort suggests that the sequential build (win a niche, then expand) is now a strategic handicap. Founders who pre-plan the full platform — even if it sounds unreasonably aggressive — are better positioned.

"As software engineering costs drop, building Act I and Act II simultaneously is no longer unrealistic."

"Ideas that look too aggressive on day one may already be too conservative by fund close."

---

Insight 2: For Agent Product Builders — Design Closed Loops, Not Open Ones

Production-viable agent products require human-designed path constraints with eval gates at each step. Open-ended agents may win demos but will fail at enterprise scale due to cost and quality unpredictability.

"A human designs the path first: clear goal, defined steps, an eval gate at each step, a handoff point. The agent loops within that structure, and each run feeds performance data into the next. Runs on a normal budget and improves over time."

---

Insight 3: For Fund Managers — Match Concentration to Fund Size or Risk Blowing Up LP Capital

The simulation data provides clear guardrails: smaller funds that over-concentrate (9-13 companies) have a 50% chance of losing LP money. Diversification is not a hedge against conviction — it's a mathematical necessity below a fund size threshold.

"Half of simulated maximum-concentration funds lose LP money."

"No construction strategy compensates for poor deal selection or insufficient ownership in the eventual outlier."

---

6. Overlooked Insights

Insight 1: The 2022 Vintage Is Already Outperforming 2021 at an Earlier Age

The article briefly notes that the 2022 vintage — only four years old — is already tracking above the 2021 cohort's final distribution. This suggests the post-correction entry prices may have quietly created a significantly more attractive vintage.

"The 2022 vintage (4 years) is already tracking better: 90th percentile 1.70x, 95th 2.02x."

This makes the 2022-2023 fund vintages potentially underappreciated by LPs still anchored to the 2021 disappointment narrative.

---

Insight 2: AI-Driven LP Reporting Reduction Could Unlock a Structurally Underserved Market Segment

Dan Gray makes a passing but significant observation: the reason LPs concentrate capital in large platform firms may be operational friction, not return-seeking. If AI reduces the overhead of managing a broader portfolio of smaller fund commitments, it could unlock a structural reallocation of LP capital toward emerging managers.

"AI-driven reductions in LP reporting friction may be the structural fix that makes a broader portfolio of smaller fund commitments economical."

This positions AI fund operations tools (like Vessel) not just as productivity plays, but as potential market-structure disruptors for how LP capital is allocated across fund sizes.