On August 9, Theory Ventures partner Tomasz Tunguz published a short 3-minute post containing hard data with major implications for AI coding toolchain valuations: Harvey, Legora, and Sierra—three AI harness (vertical industry agent platform) companies—each crossed $100M ARR within 9 months, with market valuation multiples of 50x, 56x, and 100x. The multiples themselves aren't novel—100x ARR multiples existed during the 2021 SaaS bubble—but Tunguz's key finding is that the underlying growth this time is 3x stronger. The 100x multiple is back, but the growth beneath it runs at 3x speed.
The numbers
- Harvey (legal AI agent) — announced ARR crossing $100M on Aug 4, 2025 (CEO Winston Weinberg confirmed on CNBC); $5B valuation in its Series E on Jun 23, 2025; ~$75M revenue at Series E pricing (per TechCrunch).
- Legora (legal AI agent, Europe) — announced ARR crossing $100M on Apr 2, 2026; $5.6B valuation in Series D (incl. extension), Apr 2026; revenue at pricing undisclosed.
- Sierra (customer-service AI agent, founded by Bret Taylor) — posted "Sierra hits $100m ARR milestone in 7 quarters" on Nov 21, 2025; $10B valuation in Sep 2025; revenue at pricing undisclosed, but public figure was $26M at end of 2024.
- Ramp (fintech + AI tools + card interchange revenue) — valuation basis is $1.4B revenue; its ARR multiple should be read against fintech comparables.
- Decagon (customer-service AI agent) — valuation basis is $35M revenue (end of 2024); a Jan 2026 round implies an upper-bound 129x (based on estimated Oct 2025 revenue—an estimate, not a confirmed multiple).
- Growth rate doesn't explain the multiple gaps — among the three companies past $100M ARR, the fastest grower received the lowest multiple. The market isn't paying a growth premium; it's paying a category position premium.
- Category position is the key variable — Harvey/Legora/Sierra get 50–100x because they are nearly the de facto sole players in verticals like legal AI and customer-service AI. Ramp's 85x likewise reflects its near-monopoly in fintech + AI tools. Decagon's 129x is an optimistic estimate from a still-early category—multiples compress as categories get crowded.
- Same pattern at Databricks — Tunguz notes Databricks' ARR multiple has also accelerated since January 2026, suggesting the "data + AI platform" category is being re-rated.
- 2021's 100x cycle was a product of liquidity floods and low rates — Tunguz wrote his own 100x ARR analysis on Nov 4, 2021, backed by pandemic QE and remote-work SaaS growth. Returning to 100x five years later with 3x underlying growth is a "higher-quality" 100x.
- Cloud SaaS era — 20x at $10M ARR, 10x at $100M, 5x at $1B. Multiples fall monotonically with ARR.
- AI agent era — 50x at $10M ARR, 50–100x at $100M, 30–50x at $1B. Multiples rise after crossing $100M ARR—agent platforms' network effects, data flywheels, and customer stickiness are an order of magnitude stronger than traditional SaaS.
- General AI coding tools vs. vertical agent platforms are diverging — Cursor / Devin / Replit Agent (general tools) follow SaaS logic (multiples fall with ARR); Harvey / Legora / Sierra (vertical platforms) follow network-effect logic (multiples rise with ARR). The former needs developer penetration; the latter needs industry monopoly position.
- The AI harness boundary is being redrawn — LangChain's Managed Deep Agents (Aug 8, topicId 178630987), Microsoft SkillOpt's portable experience layer (Aug 9, topicId 178603080), and Cloudflare making ADLC the SDLC default (Aug 10, topicId 178630990) all point the same way: AI harness is becoming a platform layer, not a tool layer. Tunguz's data validates this from the valuation side.
- Vertical agent platforms are the 2026 sweet spot — Harvey (legal), Legora (European legal), Sierra and Decagon (customer service), Hebbia (finance), EvenUp (legal), Phaidra (facilities), Atropos Health (healthcare)—each carries a multiple in the 50–125x range.
- 6 of Tunguz's 15 data points are estimates — 9/15 are company announcements or founder posts; the remaining 6/15 fill gaps with third-party estimates. Decagon's 129x in particular is an upper-bound estimate that will compress as the category crowds.
- Ramp's $1.4B revenue is mostly interchange — Ramp isn't pure subscription SaaS, so its multiple should be compared against fintech, not AI harness (noted in Tunguz's own footnote).
- Harvey's Series E math — $5B at ~$75M revenue implies 67x at pricing; at 50x after crossing $100M ARR, the valuation would still be $5B—a doubling of effective valuation within 9 months on the category-position premium.
- Sierra's speed — 7 quarters to $100M ARR implies ~57% geometric-average quarterly growth. What took Sierra ~21 months takes traditional SaaS companies 60–84 months on average—a 3–4x speed advantage.
- Aug 8 — LangChain Managed Deep Agents public beta (topicId 178630987) — harness as platform layer.
- Aug 9 — Prime Agent RLM Harness (topicId 178603082) — self-upgrading harness.
- Aug 10 — Cloudflare ADLC + Agents trio (topicId 178630990) — SDLC default for AI coding toolchains.
- Aug 10 — AI harness ARR multiples return to 100x (this post) — market repricing of the AI harness platform layer.
- Tomasz Tunguz original: https://www.tomtunguz.com/ai-harness-arr-multiples
- Tunguz 2021 100x ARR multiple: https://tomtunguz.com/100x-arr/
- Harvey CNBC report: https://www.cnbc.com/2025/08/04/harvey-ai-crosses-100m-arr
- Sierra $100M ARR announcement: https://sierra.ai/blog/100m-arr
- Legora $100M ARR announcement: https://www.legora.com/newsroom/100m-arr
- TechCrunch Harvey $75M revenue: https://techcrunch.com/2025/06/23/harvey-series-e
Per Tunguz, these companies' ARR multiples fall in a 25x–125x range, and have been accelerating in recent months—from January 2026, the multiples for Legora, Sierra, and Ramp all rose. This inverts the typical SaaS rule that multiples fall as ARR grows—these companies' multiples kept climbing even after passing $100M ARR.
Why the multiples differ
New pricing rules for AI agent platforms
Implications for the AI coding toolchain:
Caveats worth noting
Context in the August AI toolchain storyline
The August 10 datapoint is the market's confirmation moment: earlier August releases argued the harness is a platform; Tunguz's data proves the market agrees.
Sources: