Essays by Ry Walker
Essays on AI, startups, and software engineering by Ry Walker, founder of Tembo and Astronomer.
last 30 dayslast 90 days
On Agents
Agents Are Software
- Kill the MCP, Ship the CLI
- The Declarative Atomic Agent
- Agents Are Software, Not Prompts
- Prototype on MCP, Productionize in Code
- Sessions Replace Tasks, Runs, and Threads
- The Three Phases of AI-Assisted Engineering
- The Agent Is the Primitive, Not the Automation
- Automating Knowledge Work Is Software Engineering
- Ship the Agent Specs as Code, Not as Database Rows
Demo to Deployment
- The Mirror Problem
- The Pilot Is the Product
- The Demo Is Not the Deployment
- Start With the Mirror, Not the Model
- Vibe Code Has No Production Strategy
- Start With the Pain, Not the Platform
- The Agent Lab Replaces the Agent Deploy
- Your Agents Will Find the Dirty Data First
- The Proof of Concept Ends When Trust Begins
- Internal AI Tools Have a Twelve-Month Shelf Life
- The Operationalization Gap: Where AI Demos Go to Die
- Agents in Production: GTM Mesh and the Death of the ERP
- The First Agent You Should Ship Is the Operational Discipline Layer
Build vs Buy
- The Procurement Trap
- Homegrown Platforms Decay
- Don't Build the Eleventh CLI
- The Agent Stack Build-vs-Buy Map
- Buy the Integration Until It Breaks
- The Agent Buyer Map: Who Builds, Who Buys
- What the Build-vs-Buy Data Actually Shows
- Your Real Competitor Is the Internal Build
- The Real Product Is What Replaces Homegrown
- The Gap Is Infrastructure, Not Intelligence
- Customization Is the Moat, Not Model Quality
- The In-House Tool Dies When Its Builder Leaves
- The Reversion to Internal Software Has Started
- The Mega-Agent Fantasy Is Already Falling Apart
- Help Them In-House It, Don't Sell Them Another SaaS
- Agent Build Versus Buy: Why Engineers Keep Building It Themselves
Harness & Factory
- The Framework Trap
- The Agent Harness Problem
- The Harness Layer Has No Moat
- Software Factories Run on Pre-Prod
- The Framework Cannot Be the Product
- Your Agent Instructions Are Rotting
- The Harness Matters as Much as the Model
- Build the Infrastructure, Not the Harness
- Where the Agent Runs Is a Product Decision
- The Harness Is the Product. The Prompt Is Cheap
- The Slowest Part of Your Agent Is the Git Clone
- Agents Are Software, and Software Needs a Factory
- A Real VM Is the Primitive for Cloud Coding Agents
- The Harness Is Commoditized. Everything Else Is Not
- Give an Agent a Real Computer and It Stops Being a Coding Tool
- Investing in the Wrong Part of the Harness Is the Real Mistake
Context & Memory
- The Report Is Coming Back
- The Organizational Context Gap
- Stop Re-Deriving the Same Facts
- Agent Memory Is the Defensible Layer
- Context Engineering Is the Hard Problem
- Context Is the Moat — Don't Give It Away
- The Commit History Outlives the Veterans
- Each Person Needs Their Own Agent Instance
- Tech Context Is Tractable. Org Context Is Not
- The Real Product Is the Graph, Not the Interface
- Why Sandboxes Beat Vector RAG for Code Generation
- The Codebase Is the Territory. The Agent Needs a Map
- Tool Sprawl Is the Hidden Tax on Every Agent You Run
- Facts Rot. Reports Supersede. Build Memory Accordingly
- The Submodule Problem Is the Whole Problem in Miniature
- Three Layers of Agent Context, and Most Agents Have Zero
- Agent Memory Is Unsolved. Workflow-Scoped Learning Is Not
- Organizational Context Is the Hardest Problem Nobody Has Solved
- Two Maintenance Curves: Infrastructure Decreases, Context Never Stops
Mesh & Architecture
- The Orchestration Bet
- The Mesh, Not the Monolith
- The Convergent Agent Stack
- The Mesh of Specialists Pattern
- The Agent Infra Maturity Gradient
- Most APIs Are Not Ready for Agents
- Background Agents Are the Underbuilt Layer
- Event-Driven Agents Change What Is Possible
- The Calendar Is a Legacy Orchestration Layer
- Determinism Is a Bandage for a Bad Agent Loop
- Triggered Workflows Generate Most of the Volume
- The Orchestrator Is Where Your Tokens Go to Die
- The Bot Is the Container. The Agents Are the Work
- Start Agent Work in the Channel, Not the Terminal
- The Bot Is the Interface. The Agents Are the Units
- The Atomic Agent Mesh: Architecture, Build-vs-Buy, and the Review Layer
Workflows & Roles
- Agent Leverage Needs an Owner
- Start With Workflows, Not Roles
- With AI, Everybody Is an Engineer
- The Primitives Are the Same Across Roles
- What Workflow-First Looks Like in Practice
- One Human Will Supervise Hundreds of Agents
- Agents Are About to Break Out of Engineering
- The Operator Should Own the Agent, Not the Developer
- The Grayscale Between Engineering and Everywhere Else
- The Agent Owner Is a New Job, and It Is Not Engineering
- The Single Pane of Glass Is a Custom App, Not a Chat Window
Review & Trust
- Human Review Is Not a Limitation
- Code Review Becomes the Bottleneck
- The Algorithm Should Be Inspectable
- Review Needs a Computer, Not a Diff
- Generation Is Solved. Merging Is Not
- Review the Session, Not Just the Diff
- Inspectable Logic, Not Black Box Magic
- Controllability Beats Magic Every Time
- Review Is Not a Screen. It Is a Primitive
- Agents Start Broken and That Is the Point
- Taste Does Not Scale With Token Throughput
- Your Agent Pipeline Is Missing the Back Edge
- Non-Determinism Demands Human Correction Loops
- Merging Code Nobody Understands Is the Real Risk
- No Human Should See the PR Until It Has Been Beat Up
- The Agent Made a New Type Instead of Finding the Real One
- Review Is Not a Feature. It Is the Interface to Agent Work
- Spend Two Tokens Defending Code for Every One You Spend Writing It
- Controllability Is Not Optional. Enterprise Teams Do Not Want Magic
Coordination & Governance
- You Need a Directory of Agents
- Skills Are Software, Not Markdown
- The IKEA Effect Is Real for Agent Tooling
- The Coordination Crisis AI Tooling Created
- When Building Is Cheap, Coordination Breaks
- Users Should Iterate on Agents, Not Developers
- Agents Give the Signal Humans Are Too Polite to Send
- The Individual Contributor Picks the Model, and That Is the Cost Problem
Defense & Security
AI Coding Landscape
- The Road Ahead for AI Coding
- The Multi-Agent Platform Play
- From Code Writer to Code Judge
- The AI Coding Middleware Moment
- Democratizing Software Engineering
- Beyond the Coast-to-Midwest AI Gap
- The Cowboy Developer Era Is Ending
- The Three Tiers of AI Coding Adoption
- The Chaos in Coding Agents Is Our Friend
- Synchronous Cloud Development Is the Next Wave
- The IDE Is Going Away, and the Cloud Is Where Agents Run
- Cloud Dev Environments Are the Unlock for Parallel Agent Work
- Every CTO Is Asking How to Move Engineering Work to the Cloud
On General Business
Founder Notes
- Podcasts I Love
- Because I Had To
- Always Too Early, Never Wrong
- What I Learned Shutting Down OSSRank
- Give the Young Hire Four Jobs in Two Years
- The Underrated Power of a Positive Attitude
- Every Obvious AI Idea Gets Commoditized in Weeks
- The Weekly Progress Email Is an Artifact, Not Bureaucracy
- Remote Teams Run on Written Decisions, Not Hallway Conversations
Decision-Making & Intuition
Tembo & the AI Era
- The Tembo Manifesto
- AI Code Is Not Slop
- Why "Good Enough" Code Wins
- Personal AI Agents: The Complete Landscape
- Rise of the Agents: An AI Coding Ecosystem Map
- AI-First Software Development: Redefining How We Build Software
- I Asked 399 Developers for Their One Wish. Here's What They Said
- Measuring Developer Performance (And Why AI Might Make It Worse)
- Claude Code Just Got a Serious Upgrade, and I Can't Stop Using It
- The Future of AI Coding: Beyond Tools to True Autonomous Development
Engineering & Dev Performance
- The AI Coding Tool Wrinkle
- The GTM vs. R&D Measurement Gap
- The Gaming Problem Never Goes Away
- The 4X Mandate Without a Measuring Stick
- What Actually Helps Developer Performance
- You Cannot Measure Progress on Work You Never Defined
- Code Became Ephemeral, and Quality Became Blast Radius
- Top Performer Analysis: The Real Opportunity in AI Tool Telemetry
Open Source
Product & Growth
GTM & Platforms
- Think Small to Win Big
- Follow No Goose Absolutely
- The Design Partner Discipline
- Amplify the Tools They Already Trust
- Hourly Billing Cannot Survive Agents
- Self-Hosted Is Where the Real Pull Is
- Price for Adoption First, Profit Later
- Three Audiences, Three Agent Strategies
- Developers Want Raw Access, Not a Wrapper
- The Forward-Deployed Engineer Is the Wedge
- Catch the Buyers Before They Become Builders
- The Zero-Stickiness Problem at the Tool Layer
- The Three Services That Sell an Agent Platform
- Flexibility Is Not Optional for Agent Platforms
- From Agent to Platform: Why the GTM Is Services-Shaped
- GTM Mesh: Closing the Gap Between Ad Click and Revenue
- Wrap the Product in Services Until the Buyer Catches Up
- The Forward-Deployed Model Is the Only One That Actually Works
- The Arena of Arenas: Why There Is No Winner-Take-All in Agents
- Services Should Prepare Customers for Your Product, Not Compete With It
Markets & Strategy
- Provocatypes Over Roadmaps
- The Token Reckoning Is Coming
- The Fifth Inning and the First
- Pluck a Feather From Each Goose
- The Long Game in Agent Companies
- Everything Was an AWS Wrapper Too
- The Agent as ERP Is Not a Metaphor
- Running Out of Tokens Is a Buy Signal
- Crowded Starting Line, Empty Finish Line
- The Human Is No Longer the Integration Layer
- When Building Is Cheap, GTM Is the Only Edge
- Win at One Thing Before You Become a Platform
- The Unit of AI Consumption Is the Organization
- You Are Underspending on Tokens, Not Overspending
- The ERP Is Dead. The Agent Is Your Operating System Now
- Own the Compute Layer or Pay Someone Else's Margin Forever