Win at One Thing Before You Become a Platform
The pull to build a general-purpose agent platform is strong. But focus is a marketing asset, and you have to be undeniably the best at one workflow before you expand.
Ry Walker
Essays by Ry Walker on startups, software, and AI.
Latest posts
Where the Agent Runs Is a Product Decision
Regulated enterprises will not adopt agents that move data outside their walls. Deployment topology is a product feature, not an ops afterthought.
The Weekly Progress Email Is an Artifact, Not Bureaucracy
Informal updates evaporate. A formal weekly progress email creates a permanent, queryable record that protects both sides of a client engagement.
You Are Underspending on Tokens, Not Overspending
The instinct heading into 2027 is to optimize token spend. Most engineering teams have the opposite problem, they have not yet reached the spend that quality actually requires.
Synchronous Cloud Development Is the Next Wave
Three months ago every startup wanted async background agents. Now the leading teams want all their coding sessions in the cloud, synchronous work included.
Running Out of Tokens Is a Buy Signal
When engineering teams hit their token caps, that is not a cost problem to contain. It is the clearest signal you have that AI adoption is actually working.
Generation Is Solved. Merging Is Not
Coding agents made producing pull requests nearly free. Now the backlog lives in review, and the teams that win will engineer the review layer, not just generation.
Review Needs a Computer, Not a Diff
Agents have made code generation cheap and review the bottleneck. The fix is giving every reviewer a live machine with the change already running, not a diff in a browser tab.
The Report Is Coming Back
Non-engineering agent work needs fast, cheap context. A pile of markdown files in the same environment engineers use beats dumping everything into Slack.
Merging Code Nobody Understands Is the Real Risk
Autonomous merge sounds like speed until nobody in the org knows what the product is anymore. Human review is how a company keeps a mental model of its own software.
Developers Want Raw Access, Not a Wrapper
Teams have invested months of workflow into their coding agents. Platforms that wrap those agents in proprietary magic are asking buyers to throw that away.
Build the Infrastructure, Not the Harness
Startups cannot out-build frontier labs on models or harnesses. The durable position is the infrastructure layer that runs harnesses you do not own.
Hourly Billing Cannot Survive Agents
Professional services firms bill by the hour, so every gain in AI efficiency is revenue they destroy. Expect AI-first services firms to eat the incumbents.
Facts Rot. Reports Supersede. Build Memory Accordingly
Organizational agent memory needs rot and supersession semantics. Facts have lifespans, and the fresh report that replaces yesterday's is the format LLMs actually need.
Every CTO Is Asking How to Move Engineering Work to the Cloud
Startups are moving all of their engineering work into cloud agent environments. Enterprises are moving some. The constraint that decides how is the codebase itself.
Don't Build the Eleventh CLI
The coding CLI layer is a browser war in progress. Betting your platform on one harness is a mistake. Agnosticism is the durable position.
Get the Credentials Off the Laptop
Coding agents running in permissive modes next to production credentials on developer laptops is the quiet risk in every AI-forward engineering org. The fix is moving execution off the machine.
The Cowboy Developer Era Is Ending
The solo engineer battling code on a laptop for weeks is a dying model. Units of engineering work are becoming multiplayer, and multiplayer work has to live in the cloud.
The Commit History Outlives the Veterans
Institutional knowledge about legacy systems is walking out the door with retiring engineers. Decades of version control history is the raw material to rebuild it.
Start Agent Work in the Channel, Not the Terminal
When agent sessions start in Slack or Teams, work becomes multiplayer by default. Anyone can delegate, anyone senior can redirect, and context stops dying in tickets.
Agents Give the Signal Humans Are Too Polite to Send
The best prioritization signal in your company dies of politeness. Agents will plus-one the duplicate ticket a human never would, and that changes what engineering sees.
The Slowest Part of Your Agent Is the Git Clone
Background agents lose most of their time to environment setup, not inference. Warm VM snapshots with repos, dependencies, and skills baked in fix it.
The Agent Owner Is a New Job, and It Is Not Engineering
Enterprises need agent owners, not more prompt engineers. Maintaining a fleet of agents is mechanic work, and the people closest to the work can do it on the right fabric.
Agent Leverage Needs an Owner
Most engineering orgs have hit baseline coding agent adoption and stalled. The reason is organizational, not technical. No one owns making the agents better.
Your Agent Instructions Are Rotting
Every new model reinterprets your old agent instructions. The discipline that matters now is deletion, and almost nobody practices it.
No Human Should See the PR Until It Has Been Beat Up
As agents multiply code volume, reading diffs stops scaling. The answer is adversarial verification, where agents try to break the change before a human ever reviews it.
Spend Two Tokens Defending Code for Every One You Spend Writing It
Most teams pour AI budget into new features while their existing code rots. The ratio should invert: agents should defend critical code, not stay away from it.
The Three Services That Sell an Agent Platform
The path to enterprise agent adoption runs through three distinct service engagements, and each one exists to pull the customer deeper into the product.
Stop Re-Deriving the Same Facts
The practical win in agent memory is not total recall. It is a shared fact store with a lifecycle, so agents stop re-deriving what the organization already knows.
Services Should Prepare Customers for Your Product, Not Compete With It
The services you sell should get customers ready to succeed with your product, not build one-off agents that pull you away from it.
Software Factories Run on Pre-Prod
Autonomous agent loops are real, but almost every impressive self-driving software factory is running on code without customers. Production is a different game.
Review the Session, Not Just the Diff
When an agent writes the code, the diff is half the artifact. The session that produced it carries the intent, and reviewers need access to it.
Review Is Not a Feature. It Is the Interface to Agent Work
The unglamorous work of making agent review legible (request changes, jump-to-context, re-request review) is where enterprise agent deployment actually lives.
Price for Adoption First, Profit Later
Your first ten agent engagements should be priced to win, not to profit, because you are buying learning and reference cases across platforms you have never touched.
The IDE Is Going Away, and the Cloud Is Where Agents Run
Local machines cannot run ten agents at once. The cloud dev environment stops being a nice-to-have the moment agentic coding becomes the default.
The Real Product Is the Graph, Not the Interface
When your polished front-end is a byproduct of an underlying knowledge graph, be honest about which layer is the product and which is the tech demo.
The Fifth Inning and the First
San Francisco is mid-game on agentic engineering while most enterprises are just starting. That diffusion lag is not a problem. It is the market.
With AI, Everybody Is an Engineer
The SREs who never wanted to write code no longer get a choice. When AI turns intent into working software, the line between engineer and non-engineer disappears.
Remote Teams Run on Written Decisions, Not Hallway Conversations
When your team spans time zones, the decisions made in the room while others sleep have to be written down or they never happened for half the company.
Code Became Ephemeral, and Quality Became Blast Radius
Code has always recycled itself, but AI collapsed the timeline. The winning move is ring-fencing the code where a bug is catastrophic and letting the rest move fast.
Determinism Is a Bandage for a Bad Agent Loop
Enterprise teams reach for deterministic workflow builders when their agents fail. The real fix is better tools the agent can actually see and call.
The Calendar Is a Legacy Orchestration Layer
Most enterprise coordination still runs on meetings because humans are the scheduler. Background agents replace synchronous huddles with continuous execution.
Awareness Is the Hardest Problem, and Content Is the Secret Weapon
For technical products, the bottleneck is rarely closing deals or fixing the product. It is awareness, and the answer is founder-led content, not more ad spend.
Give an Agent a Real Computer and It Stops Being a Coding Tool
When you give an agent a full VM, real tools, and an easy way to talk to it, the coding tool quietly becomes a general knowledge-work platform.
Investing in the Wrong Part of the Harness Is the Real Mistake
Not every agent product needs a vector database and a custom retrieval pipeline. The hard part is knowing which layer of the harness deserves real investment.
Wrap the Product in Services Until the Buyer Catches Up
Non-technical buyers cannot configure their own agents yet. The fastest path to revenue is selling the service of building agents for them.
A Real VM Is the Primitive for Cloud Coding Agents
Containers cannot replicate a developer's local environment. Moving coding agents to the cloud requires a full virtual machine, and that is harder than it looks.
Tool Sprawl Is the Hidden Tax on Every Agent You Run
Why exposing an agent to a thousand MCP tools quietly multiplies cost and risk, and why locking tools down per agent is the operational discipline that actually works.
Self-Hosted Is Where the Real Pull Is
Customers rarely ask for self-hosted, but it is where serious enterprise pull lives, where margins are clean, and where stickiness compounds with every separate deployment.
The Proof of Concept Ends When Trust Begins
The real signal a pilot is working is not a feature checklist. It is the customer saying they are comfortable starting the vendor process. That moment is earned, not pitched.