Writing

Thinking out loud about the stack.

Essays and notes from Marcus Huang and Natalie Ferris on AI agents, LLM infrastructure, and what we're learning from the portfolio.

May 2026

The Compound Agent

Single-step agents were the proof of concept. What comes next is agents that delegate, spawn subagents, and compose — the compounding logic that makes autonomous systems genuinely useful.

Marcus Huang agents

Mar 2026

What Agentic UI Actually Means

The interface layer for AI agents is not a chatbot with better styling. It is a fundamentally different design problem — one that most product teams have not yet started to solve.

Natalie Ferris product

Jan 2026

Multi-Agent Systems Are Not a Framework Problem

The conversation about multi-agent coordination keeps returning to orchestration frameworks. That is the wrong level of abstraction. The hard problems are about state, trust, and failure modes.

Marcus Huang infrastructure

Nov 2025

Memory as Infrastructure

Agent memory is not a feature you add to a chatbot. It is a distinct infrastructure primitive — with its own consistency models, retrieval semantics, and failure profiles. We are still early in understanding what it means to build this well.

Natalie Ferris infrastructure

Sep 2025

The Prompt Is Not the Product

A prompt is a starting configuration, not a moat. The companies that will matter in the AI application layer are the ones building infrastructure around prompts — versioning, evaluation, deployment, observability.

Marcus Huang developer tooling

Jul 2025

Evaluating AI Agent Reliability

Reliability for LLM applications is not a boolean property. It is a distribution over inputs, contexts, and time. Building evaluation systems that reflect this reality is one of the genuinely hard unsolved problems in production AI.

Natalie Ferris evaluation

Jun 2025

Why We Backed Letta

The case for stateful agent memory as a standalone infrastructure category — and why the MemGPT lineage represents something more durable than a research project.

Marcus Huang portfolio

Apr 2025

The Tool-Use Gap

LLMs can call tools. That was the easy part. The hard part is building the execution layer that makes tool-calling reliable, observable, and recoverable when things go wrong — which they will.

Natalie Ferris agents

Feb 2025

LLM Costs Are Falling. Infrastructure Margins Are Not.

Token prices keep dropping. But the hard costs of running AI in production — latency management, context assembly, observability, eval — are not model costs. They are engineering costs. Those are not falling.

Marcus Huang infrastructure

Jan 2025

Predictions for the Agentic Stack in 2025

What we expect to see consolidate, what will remain fragmented, and where the most interesting new companies will emerge in the year ahead.

Natalie Ferris outlook

Nov 2024

Context Windows Changed Everything. Memory Still Matters.

Longer context is not a substitute for persistent memory. The architectural tradeoffs are different, the use cases are different, and the infrastructure requirements are different. Both will matter.

Marcus Huang infrastructure

Sep 2024

Developer Tooling for the AI Era

The tools developers need to build AI-native applications are not the same tools they used to build cloud-native applications. We are watching an entirely new developer toolchain emerge in real time.

Natalie Ferris developer tooling

Aug 2024

The RAG Taxonomy

Retrieval-augmented generation is not one thing. Naive RAG, advanced RAG, modular RAG, agentic RAG — each is a distinct architectural pattern with different performance characteristics and infrastructure requirements.

Marcus Huang infrastructure

Jun 2024

Orchestration Is the New Middleware

Every generation of software has its coordination layer — the middleware that wires together the parts that were not designed to work together. For AI-native applications, that layer is LLM orchestration.

Natalie Ferris infrastructure

Apr 2024

Why Seed Stage Still Matters in the LLM Era

The AI hype cycle has compressed Series A timelines and inflated valuations at the growth stage. That makes precise seed-stage investing more valuable, not less — if you can stay disciplined about what you are actually underwriting.

Marcus Huang investing

Feb 2024

The Second Wave of LLM Applications

The first wave was about adding a chat interface to everything. The second wave will be about designing products that are AI-native from the ground up — different distribution, different architecture, different retention dynamics.

Natalie Ferris product

Jan 2024

Closing Fund II

We closed Hearthstone Fund II at $47M. Some thoughts on what has changed since Fund I, what the portfolio taught us, and where we are pointing Fund II.

Marcus Huang firm

Nov 2023

What Production AI Actually Looks Like

The gap between AI demos and production AI systems is larger than most people realize. Latency budgets, context management, fallback logic, evaluation pipelines — this is the engineering work that never shows up in launch announcements.

Natalie Ferris infrastructure

Oct 2023

The Agent Runtime Problem

Agents need a runtime — a substrate that handles execution, interruption, resumption, and state persistence across long-running tasks. Nobody has fully solved this yet. The solutions that emerge will be infrastructure companies.

Marcus Huang agents

Aug 2023

LLM Observability Is Underrated

When an LLM application fails in production, you need to know why. The observability tools that exist for conventional software are not adequate for LLM pipelines. This is a large unsolved problem that most teams are trying to solve with logs.

Natalie Ferris developer tooling

Jun 2023

Why We Invested in Greptile

Code understanding is one of the most tractable near-term applications of LLMs — and one of the most valuable. When a model can hold a full repository in mind, code review, search, and documentation become entirely different problems.

Marcus Huang portfolio

Apr 2023

ChatGPT and What Comes After

The ChatGPT moment was real and it accelerated everything we expected. What it did not change: the structural argument for infrastructure over applications. If anything, it made that argument stronger.

Natalie Ferris outlook

Nov 2022

The State of AI Tooling

Two years into Hearthstone, a survey of what the AI developer tooling ecosystem looks like, where the gaps remain, and what we expect the next generation of infrastructure companies to look like.

Marcus Huang developer tooling

Mar 2022

Foundation Models as Platform

Foundation models are not finished products. They are platforms — and like all platforms, they create surface area for new companies to build on. The question is which parts of that surface area represent durable businesses.

Natalie Ferris outlook

Sep 2021

Why Infrastructure Before Application

A year into Hearthstone, an attempt to articulate the underlying logic of the thesis: why we believe the infrastructure layer for AI will produce the most durable companies of this cycle, and why that window is open now.

Marcus Huang thesis