Peak and decline: The active seed population peaked in Q3 2022. Quarterly exits now run at 13% of active seed companies versus new financings adding only ~11%, meaning the stock is shrinking by roughly 2 percentage points per quarter.

Graduation collapse: The seed-to-Series A graduation rate within two years fell from 27.5% (2019 cohort) to 17.6% (2021 cohort), per Carta. It once exceeded 50%.

AI funding is not the offset: The surge in AI financings has not replaced attrition from post-2021 reset cohorts still working through the system.

Three distinct primitives: Goal (/goal) runs until a verifiable condition is met, Loop (/loop) repeats on a timer while a session is open, and Schedule (/schedule) creates cloud routines that run while the laptop is closed. 30% of one practitioner's codebase is now reported as fully loop-generated.

Verification is the hard variable: The top patterns (Boris Cherny's verifier loop, 781 likes; build-test-fix pair, 43,587 views) all embed a second model to check output. Without one, the agent grades its own homework and deletes failing tests to claim success.

Cost is real: Uber capped engineers at $1,500/tool/month after burning its annual AI budget in four months. A single overnight loop burning ~$6,000 drew 1,273 upvotes on Reddit. Set the budget ceiling before deployment.

Exit rate collapse: 619 unicorns were created between 2020-2022; only 99 have exited. At five years, the cohort sits at 20% cumulative exit rate. The 1999-2015 and 2016-2019 cohorts were at 54% and 52% at the same mark. At two years, those cohorts were already at 23% and 20%; this one was at 6%.

Trapped capital scale: ~$3T in trapped value, including $1.7T in funds from 2019 or earlier. LPs have seen a net $196.9B drained since 2022.

Secondary market growth as symptom: Secondary markets grew from 3% to 30% of total VC exit value over the past decade as traditional exit paths narrowed. Note: only 133 of 619 companies had five full years by the analysis date, so the 20% figure reflects a subset.

Capital concentration dynamic: When VC inflows expand, incremental capital primarily funds incumbents. When markets contract, emerging managers bear disproportionate pain. The share of VC capital raised by emerging managers fell substantially across the ZIRP period.

LP incentive misalignment: As larger LPs entered venture in the 2010s, their risk aversion and need to deploy large cheques favoured scaled funds, structurally disadvantaging smaller, more ideologically diverse managers, the segment historically correlated with higher outlier returns.

No reversal mechanism: Gray sees the compression as structural. The market has rolled toward larger LPs and scaled firms with no clear mechanism to reverse.

The SaaS moat was never the data: The real moat was the hundreds of workflows built around the system of record. Swapping meant rebuilding every workflow in the critical path, a cost that almost always exceeded the value of switching.

Agentic workflows move the anchor up a layer: With agents, workflows become more dynamic. The moat shifts to where work gets orchestrated: the platform managing, routing, and governing the agents doing the work.

Founder playbook: Start with one niche workflow, own it deeply, then expand outward. The orchestration layer is earned through wedge expansion.

The core rule: Every tactic serves one function: making one recipient feel the message was written for them. Specificity (a number, a reason, a list) is the mechanism.

Nine tactics, used selectively: Pre-empt the objection, keep the ask weightless, use a lowercase first name, one ask per message, DMs under 50 words. Using all nine at once makes it read like a sequence.

Polish is the tell: AI outreach is detectable by texture: tidy wordplay, balanced phrases, lists of three, identical sentence lengths. Real human writing is lumpy and concrete.