Third Vector Observer | Edition 3

Welcome to the Third Vector Observer, an easy-to-digest newsletter featuring 3-5 high-quality, insightful pieces on agentic AI and the impact on organizations and markets. Delivered to your inbox roughly once a month.

Things We Observed...

1. Your Data Agents Need Context (and a Semantic Layer Won’t Cut It)

a16z published a piece that crystallizes something we think is fundamental: data agents keep failing because nobody gave them the business context they need to do anything useful.

The example is deceptively simple. Ask a data agent “what was revenue growth last quarter?” and watch it drown. Which revenue definition? ARR or run rate? Which fiscal quarter? Which table is the source of truth when finance uses fct_revenue but the data team built mv_revenue_monthly? The model can write flawless SQL and still return garbage.

Semantic Layer vs. Context Layer vs. Ontology

These three concepts keep getting conflated, but they’re distinct:

• Semantic layer (LookML, dbt metrics): Defines specific metrics in code. Revenue equals X, churn equals Y. Useful but narrow, mapping well to BI dashboards where queries are predictable.

• Context layer: Broader. Encompasses metric definitions but adds entity relationships, identity resolution, tribal knowledge, governance rules, and workflow-specific instructions. Think of it as the difference between giving someone a glossary versus six months of institutional knowledge.

• Ontology: Formally defines how entities relate to each other. A customer has accounts, accounts have subscriptions, subscriptions generate revenue events. Palantir built a massive business on exactly this idea.

The ontology is the structural backbone; the context layer is the intelligence that makes it operational for agents.

Where the Market Is Heading

a16z maps out three categories of emerging solutions:

• Data gravity platforms (Databricks, Snowflake) adding lightweight context features

• Existing AI analyst companies pivoting to context-first approaches

• Dedicated context layer startups building from scratch

Encoding every business rule into YAML files by hand won’t scale. We believe the winners will crack automated context construction with human refinement.

Your Data Agents Need Context

2. Anthropic’s Revenue Growth Is Breaking the Playbook

Full disclosure: we at Third Vector are massive fans of Claude and have integrated Claude Cowork and Claude Code into how we work. So we’re biased. But the numbers Tomasz Tunguz laid out this week are staggering on their own.

Anthropic added $10 billion in revenue in a single month. That’s twice Databricks’ annual run rate, absorbed in thirty days.

The Trajectory in Perspective

• ServiceNow: 20 years to hit $10B annual revenue

• Shopify: 18 years

• Palo Alto Networks: 19 years

• Anthropic: 42 months

Can Anthropic Surpass NVIDIA?

Tunguz’s core question is when, not whether. NVIDIA currently generates $215B in annual revenue at a 22x multiple, producing a $4.8 trillion market cap. At a 25x forward multiple, Anthropic would need roughly $200B in annual revenue to match. The bull case says three years. The base case says four. Even the bear case gets there in seven.

The critical caveat is customer concentration risk. But the sheer velocity is unlike anything the software industry has produced.

What strikes us is what this means for the broader ecosystem. When an AI model company grows this fast, every adjacent category gets pulled forward: infrastructure, tooling, implementation services, data preparation. The companies building agentic workflows on top of these models are riding actual exponential demand.

When Will Anthropic Surpass NVIDIA?

3. SAP Buys Reltio: Do We See A Pattern Emerging?

Salesforce acquired Informatica. Now SAP is buying Reltio. Two of the largest enterprise software companies on the planet looked at their agentic AI ambitions and came to the same conclusion: the data underneath is a mess, and they can’t fix it themselves.

What Reltio Does

Reltio is a cloud-native master data management (MDM) platform that creates a single, consistent view of customers, products, suppliers, and employees across both SAP and non-SAP systems. That last part matters.

SAP’s head of product engineering, Muhammad Alam, was quoted saying:

“AI cannot reach its full potential when data is fragmented across business units, platforms and domains.”

In other words: if your agents can’t trust the data, they can’t do anything useful. SAP launched its Business Data Cloud a year ago, but building clean data infrastructure from scratch is slow. Buying Reltio gives them entity resolution and data harmonization that would have taken years to build internally.

The Bigger Signal

ERP companies are waking up to the fact that decades of bolt-on acquisitions and customization have left enterprise data fundamentally incompatible with autonomous AI workflows. Agentic systems need consistent entities, resolved identities, and governed data. That’s precisely what MDM companies have been building for years, mostly as unsexy infrastructure nobody wanted to pay for. Now it’s suddenly strategic.

SAP Pumps Up Agentic AI and Data Chops with Reltio Acquisition

4. Claude Managed Agents: The Infrastructure Layer Catches Up

Until now, building production-grade AI agents meant stitching together your own sandboxing, authentication, error recovery, and orchestration. Months of plumbing before a single line of agent logic. Anthropic just collapsed that timeline.

Claude Managed Agents, launched April 8, is a suite of composable APIs that handles the operational complexity of deploying cloud-hosted agents at scale:

• Secure sandboxing for code execution

• Built-in credential management with scoped permissions

• Persistent checkpointing that survives disconnections

• End-to-end tracing so you can inspect every tool call and decision point

In internal testing, the platform improved task success rates by up to 10 points over standard prompting loops, with the biggest gains on complex, multi-step tasks.

Enterprise Adoption Signals

The early deployments are production, not proofs of concept:

• Notion is using it for parallel task execution agents

• Rakuten deployed agents across product, sales, marketing, and finance within weeks

• Sentry built a debugging agent that diagnoses issues and writes patches

Also notable: the multi-agent coordination capability (in research preview) allows agents to delegate subtasks to other agents in parallel. That’s the architecture pattern everyone’s been talking about, now available as a managed service.

What This Changes

The build-versus-buy calculus just shifted. Security and compliance overhead alone used to be a six-month project. When the model provider handles sandboxing, identity, and tracing natively, the barrier to shipping drops to the quality of your prompts and tool definitions. We’re moving from “can we build this?” to “how fast can we ship it?”

Claude Managed Agents: Get to Production 10x Faster

5. 1,300 Marketing Skills Walk Into an AI Agent

Buron just dropped something that caught our eye: a library of over 1,300 pre-built marketing skills designed to plug directly into Claude and other AI agents. Each skill is a structured knowledge guide covering a specific marketing task, from diagnosing landing page conversion problems to running SEO audits optimized for AI engine discovery.

What’s in the Box

The skills span six categories:

• SEO, including audits, schema markup, E-E-A-T authority building

• Paid Ads, including Google Ads account analysis and conversion tracking

• Content Marketing, including creation, repurposing, optimization

• Strategy, including campaign planning and competitive analysis

• Analytics, including performance diagnostics and attribution

• Social Media, including campaign planning and execution

There’s a community leaderboard where skills get rated, which creates a natural quality filter. The top-rated ones read more like specialist playbooks than simple prompt templates.

Why This Matters Beyond Marketing

We’ve been building custom skills at Third Vector, and the craft involved in writing a genuinely useful skill is non-trivial. You need to encode domain expertise, edge cases, and decision logic. Buron is betting that a marketplace approach can scale this faster than every company building bespoke.

Buron Marketing Skills

Subscribe now

Welcome to the Third Vector Observer, an easy-to-digest newsletter featuring 3–5 high-quality, insightful pieces on venture operations, agentic AI, finance, and Europe's tech ecosystem. Delivered to your inbox roughly twice a month.

Things We Observed…

1. Claude Cowork Triggers $285B Selloff Across Software, Financial Services, and Asset Management Stocks

Anthropic’s release of Claude Cowork spooked markets in January, but it was last week’s launch of 11 open-source plugins that turned a tremor into an earthquake. By February 3rd, a massive $285B selloff hit software, financial services, and asset management stocks globally.

Why the panic?

The market has reached a tipping point. Investors are no longer viewing AI as a "feature" for existing SaaS; they are viewing it as a replacement for the application layer. When Anthropic—an infrastructure provider—released ready-made vertical solutions for legal, finance, and sales, the platform officially became the competitor.

What makes Cowork different: agentic autonomy

Claude Cowork doesn't just chat; it executes multi-step workflows directly on a user’s machine.

Cowork includes a library of plugins for common knowledge work functions:

• Productivity — Manage tasks, calendars, and daily workflows

• Enterprise search — Find information across your company’s tools and docs

• Sales — Research prospects, prep deals, and follow your sales process

• Finance — Analyze financials, build models, and track key metrics

• Data — Query, visualize, and interpret datasets

• Legal — Review documents, flag risks, and track compliance

• Marketing — Draft content, plan campaigns, and manage launches

• Customer support — Triage issues, draft responses, and surface solutions

• Product management — Write specs, prioritize roadmaps, and track progress

• Biology research — Search literature, analyze results, and plan experiments

“Your Margin is My Opportunity”

The damage was swift and targeted at companies whose business models rely on “seat-based” knowledge work:

• Legal Tech: Thomson Reuters (-16%), RELX/LexisNexis (-14%), and Wolters Kluwer (-13%) saw massive wipes.

• Enterprise SaaS: Heavyweights like Salesforce and ServiceNow fell 7%, as the threat of “seat compression” became real.

• Outsourcing: Indian IT giants (Infosys, Wipro) slid 5-6% as investors feared agentic AI would cannibalize FTE-based contracts.

Software ate the world, now AI is eating software. As Jeff Bezos famously said, “Your margin is my opportunity.” In 2026, AI isn't just eating software; it's eating those fat, juicy SaaS margins. Yum.

2. Dead Companies Walking - Lessons from a Short-Seller Applied to SaaS

In this sobering piece, Tomasz Tunguz applies the “six failure modes” of legendary short-seller Scott Fearon (Dead Companies Walking) to the current state of the software industry. Tunguz’s thesis is clear: the “hairpin turn” from SaaS to AI is amplifying every classic mistake that causes companies to zero out.

Fearon’s Six Failure Modes (Applied to the AI Era):

1. Learning only from the recent past: SaaS multiples have been flat for three years. Relying on 2021-era growth playbooks is the fastest way to run out of runway.

2. Relying too heavily on a formula: The “efficient growth” formulas of the last decade (PLG vs. Enterprise) are breaking. In an AI-native world, sales motions and quotas must be rebuilt from scratch.

3. Misreading or alienating customers: Customers don’t want “SaaS with an AI wrapper”; they want AI-native solutions that solve pain points in fundamentally new ways. Selling the old application leads to instant churn.

4. Falling victim to a mania: Shipping half-baked AI features to appease a roadmap isn’t the same as delivering value. As Tunguz notes, 2026 is the year where “long-running agents” will separate the wheat from the chaff.

5. Failing to adapt to tectonic shifts: The shift from seat-based pricing to outcome-based or compute-based models is a structural earthquake.

6. Being physically or emotionally removed: If leadership isn’t using tools like Claude or GPT daily, they aren’t just behind—they are “emotionally removed” from the reality of the technology.

Link: https://tomtunguz.com/dead-companies-walking/

3. The High-Growth Handbook: Molly Graham on Lenny’s Podcast

In this episode of Lenny’s Podcast, Molly Graham (Glue Club, ex-Facebook/Google) shared a masterclass on leading through the chaos of hyper-growth. For anyone building or advising ventures, her frameworks and mental models help moving from “scrappy startup” to “scaled machine.” See selected highlights below the Spotify link.

The Waterline Model

Molly shared a useful framework, the “Waterline Model”, for diagnosing team dysfunction. She argues that when a team is struggling, leaders often default to “scuba diving”—going deep into interpersonal conflicts or individual performance issues. Instead, she suggests you should “snorkel” first.

• Above the Waterline: These are structural issues—unclear roles, lack of alignment on goals, or poor processes.

• Below the Waterline: These are the messy, interpersonal “people” problems.

The Insight: 80% of what we perceive as “personality clashes” or “bad hires” are actually just structural failures above the waterline. Before you blame the person, fix the role clarity and the goal-setting. It’s a powerful reminder for venture operators: Structure dictates behaviour.

Stop Over-Investing in Low Performance

Molly’s advice centers on the opportunity cost of a manager’s time. She argues that the common instinct—spending 80% of your energy on the bottom 10%—is a strategic error.

“The mistake most managers make is they spend all their time on their lowest performers... it’s actually a disservice to the rest of the team. You have to get comfortable with the fact that if it’s not working, the kindest thing you can do is move them out quickly. Then, take all that energy and pour it into your ‘high-potentials’—the people who are actually going to build the future of the company.”

Other Key Takeaways:

• “Give Away Your Legos”: The classic framework for scaling yourself. If you aren’t actively handing off the parts of your job you love, you’re becoming the bottleneck.

• The “J-Curve” Career: Why the most successful paths often look like a series of terrifying leaps into the unknown rather than a steady climb up a corporate ladder.

• Strategy Should Hurt: If your strategy doesn’t involve making painful trade-offs (saying “no” to good ideas), it’s not a strategy; it’s a wishlist.

Subscribe now

Welcome to the first edition of Third Vector Observer, an easy-to-digest newsletter featuring 5–10 high-quality, insightful pieces on venture building, AI, and news from Europe's tech ecosystem. Delivered to your inbox roughly twice a month.

Things We Observed…

1. The Economist: AI Adoption

A reality check on the state of AI adoption—and a reminder that not everyone is leaning in just yet.

“Adoption has fallen sharply at the largest businesses, those employing over 250 people. Three years into the generative-AI wave, demand for the technology looks surprisingly flimsy.”

And:

“From today until 2030 big tech firms will spend $5trn on infrastructure to supply AI services. To make those investments worthwhile, they need on the order of $650bn a year in AI revenues, according to JP Morgan Chase, a bank, up from about $50bn a year today.”

The Economist - Investors expect AI use to soar, that's not happening (possibly behind a paywall)

2. Fred Wilson: Making AI Virtual Advisors

I’ve been reading Fred Wilson’s blog, AVC, for over 15 years. As a founding partner of Union Square Ventures, Fred pioneered blogging in the VC industry, sharing insights on tech trends, the state of VC, and practical advice for entrepreneurs. He maintains a massive following, even after moving his blog “onchain” (to the blockchain).

When I started working with startups ~10 years ago, I benefited from his frameworks on cap table management, startup valuation, and emerging tech like machine learning.

In this post, he discusses creating AI virtual advisors. It resonates because I’ve created a virtual CFO, CMO, and CTO using Google’s Gemini (Gems) for Third Vector. Interacting with these advisors every day is a true productivity multiplier.

AVC - Making Advisors

3. GTM in the AI Era

Lenny’s Podcast is a pure treasure trove. Lenny is the chillest, most humble host in the space. This episode with Vercel’s COO Jeanne DeWitt Grosser (formerly of Stripe and Google) on what GTM looks like in 2026 is packed with insights.

To pick just one, in three simple steps Jeanne’s team built a “Lostbot” to learn from lost deals:

Step 1: Pull a list of lost deals from the previous quarter by size.
Step 2: Gather all sales transcripts (Vercel uses Gong), Slack convos, and email threads.
Step3: Use an AI agent to diagnose each deal and extract learnings.

They then used those lessons to create a “Dealbot” that analyzes live deals to flag risks—like, “Hey, you haven’t talked to the economic buyer yet.” Essentially, these bots help Vercel find bugs in their go-to-market process.

4. AI dictation with Typeless

AI-native voice-typing got really, really good. It still feels a bit weird to talk to your computer, but it’s incredibly fast (4x faster than typing) and the transcription quality is stunning. I’ve been using Typeless for note-taking in Obsidian and for drafting emails; I love it. They offer a free tier that is well worth checking out.

5. Gaming and AI Research Lab with Dutch roots Raises $133.7M in Seed

It turns out video game recordings are a gold mine for training AI on scenarios that are expensive to simulate—like helicopter crashes or plane accidents. Pim de Witte, the Dutch founder of Medal.tv (a platform that beats YouTube in annual video game clip uploads), decided to spin off a dedicated research lab for foundation models trained on gaming footage.

The lab, called General Intuition, raised $133.7M in seed funding to build models for use cases requiring spatial and temporal reasoning, such as autonomous vehicles and robotics. While now headquartered in NYC, General Intuition’s parent company is based in Naarden, Netherlands.

6. Protocols and Agentic AI Foundation

This article from The Economist (Dec 13, 2025) does an excellent job explaining how different AI protocols, specifically A2A and MCP, work together to create the next version of the web.

The Economist - The next version of the web will be built for machines not humans