Thinking in Public
Essays, frameworks, and observations on context architecture, AI product strategy, and building what matters.
The AI Software Market Map: Pick Your Floor
Four companies you already pay for just made the same move. The AI software stack is sorting into five floors. Here is the map, and how to pick yours.
If an AI Harness Decides When Your Product Is Useful, Does It Own You?
a16z says AI agents are pushing enterprise software headless. The cannibalization risk SaaS companies are underestimating: giving up not the screen, but the decision about when your product is useful.
Knowledge Agents vs Task Agents: Scoping Agent Identity
Two kinds of agents, knowledge and task, and why scoping an agent's identity is becoming core to the PRD. A product leader's take on designing a team of agents.
Breaking Model Dependency: LLM Failover for AI Products
A major model outage exposed the cost of hardcoding one LLM. Why single-model dependency is a retention risk, and three ways to add failover to your AI product.
Context Layer vs RAG: Retrieval Is a Tactic, Not an Architecture
RAG is a retrieval pattern. A context layer is the architecture around it. The difference is why your pipeline can fetch the right chunks and still ship the wrong answer.
You Can Clone the Agent. You Can't Clone the Taste.
Salesforce paid $3.6B for Fin, the company that used to be Intercom. Read the headline and you'll assume it bought agent technology. It didn't — it already has Agentforce and frontier models. Benioff led with the team. That ordering is the whole story: the model is the commodity now, and the judgment about what to build with it is the part you can't download.
The Birthplace Test: Every Method Is Downstream of the Org That Birthed It
In 2025 a startup co-founder told me we were adopting Shape Up. The whole evaluation: his YC batchmates swore by it. There wasn't a second name on the list. Shape Up's load-bearing condition isn't the six-week cycle — it's leadership that can make a bet and leave it alone for six weeks. He couldn't leave one alone for two days. Every method is downstream of the org that birthed it.
The models are getting smaller. The context layer's job just got bigger.
Small language models with the right context are starting to deliver real utility. WWDC showed the models. The context layer that feeds them is the open question.
Pick Your Floor: One Question Per Floor
The AI software market is sorting into five floors, from raw inference up to agent-native data. Naming them is easy. Holding one is not. Each floor has exactly one question that decides whether you can actually hold a position there — and most teams answer for the floor they wish they were on, not the one they're standing on.
Five Engineering Moves, Five Architectural Decisions
An engineer once asked me whether we should add a caching layer to cut latency. Good question, wrong altitude — we hadn't yet decided which context should persist at all. Lance Martin's context-engineering taxonomy gives engineers five moves on a context window. Each one has an architectural decision that should come first. The move is reversible in an afternoon; the decision shapes the product for a year.
SaaS Isn't Dying. It's Being Demoted.
Ramp's data says SaaS is fine. Notion, Linear, and Dropbox say otherwise. Why SaaS is being demoted from interface to backend, and where premium pricing now compounds.
Why “No” Doesn't Work — and What to Say Instead
Half a dozen PMs brought me the same problem this year, every one framed as “how do I tell my VP no.” Wrong question. Saying no doesn't work the way we think it does — psychological reactance guarantees the request just reroutes. The move that's served me through three startups and IBM's leadership program: scope toward the outcome, not against the request.
Why the Context Layer Loop Looks Like the Brain
The five-step Context Layer loop maps onto how memory works in the brain. Four parallels, where practitioners are converging, and three concrete fixes if your retrieval-tuning isn't fixing the actual problem.
n=3 Again: The Consumer AI Inflection Point
In 2010 I watched three strangers at a bar scroll instead of talk, and knew social media had crossed from early adopters to everyone. I had the same n=3 moment about AI this week — and the signal that matters isn't the founder shipping an agent. It's the former physician in Ecuador using ChatGPT every day.
The Composition of an Agent Team
"How many agents?" is the wrong starting question. The four-layer framework I used to compose a 14-agent software factory, with the count as the residue.
What a Context Layer Actually Is (And Why Agent Memory Isn't One)
Agent memory got the headlines. The Context Layer didn't. The four-layer separation practitioners are converging on, what actually sits inside the Context Layer, and why architecture defines infrastructure.
Market Context Wins the Deal. Situated Context Builds the Moat.
The AI industry has reduced context to what fits in the window. That's a category error. Context is a property of organizations, not model architecture — and only one of its two forms becomes a moat.
How I Built a 14-Agent Software Factory on a Single VPS
While on paternity leave I stood up fourteen agents on a single $50/mo VPS, watched three vendor moves break my plan in seven days, and learned that the infrastructure is not the moat.
Data Is Not Context
The AI industry is obsessed with retrieval — how to get the right chunks into the context window. But retrieval is downstream. If what you're retrieving was never curated, synthesized, consolidated, prioritized, or stored intelligently, your RAG pipeline is just efficiently delivering noise. This essay defines the five-step architecture that turns raw data into decision-ready context.
The Context Layer
A newsletter on the architecture behind AI memory, context synthesis, and building products that think. Launching Q3 2026.
Want to discuss these ideas?
Get in touch →