Every company building in the AI agent space right now sees the same future. Agents will be software. They will run in the background. They will handle an expanding share of knowledge work. The controversial question is how you build a business on top of that future without ending up as free R&D for the big labs — and underneath that, a harder one: how do you stay in the fight long enough to find out? The race is a marathon you sprint through. Frameworks commoditize. Monetization pins to infrastructure. And the moat is proximity to the people who are actually building. I broke this argument into ten atomic posts. Read them in any order:
- Always Too Early, Never Wrong — The compulsion that makes serial founders show up before the wave forms.
- Crowded Starting Line, Empty Finish Line — Most entrants in hot categories pivot or die. The race goes to who keeps running.
- The Framework Trap — LangChain pivoted to LangSmith. E2B sells the sandbox. The harness is the giveaway.
- What the Build-vs-Buy Data Actually Shows — The middle of the stack is built in-house everywhere. That is the opening.
- The Organizational Context Gap — The one universal unsolved layer. Whoever cracks it captures the next floor.
- Agents Are Software, and Software Needs a Factory — The harness is a wrapper. The factory is the platform.
- The Orchestration Bet — The best agent changes every quarter. The orchestrator outlasts them all.
- The Zero-Stickiness Problem — Six months of configuration ditched overnight. Why tool-layer positions are fragile.
- Pluck a Feather From Each Goose — The design partner discipline that prevents building three products at once.
- Provocatypes Over Roadmaps — When the product shape is unknown, deliberately weird prototypes beat planning.
The technology will not wait for the market to catch up. The market will not wait for any single company to win. The companies that survive long enough to matter are the ones that build for the now, monetize the infrastructure underneath, and stay close enough to real builders to know what to ship next.
— Ry
Related Essays
Crowded Starting Line, Empty Finish Line
The AI coding agent space looks crowded today. The vast majority of entrants will pivot, run out of money, or chase the next shiny thing. The race goes to who keeps running.
Pluck a Feather From Each Goose
Following every engaged customer ends with three products instead of one. The discipline is to pluck a feather from each passing goose and follow none absolutely.
Provocatypes Over Roadmaps
A provocatype is a prototype built to provoke conversation, not to ship. When the product shape is genuinely unknown, provocatypes surface assumptions that requirements gathering misses.
Key takeaways
- The startup race goes to those who keep running longest, not those who sprint hardest in the first mile.
- Agent frameworks are commoditizing fast — durable monetization pins to infrastructure layers like sandboxing, observability, and orchestration.
- Building for the now beats building for the future. Whoever captures the present market is best positioned to capture what comes next.
- Revenue per employee is the defining metric of the AI era, not headcount.
- Enterprise adoption closes in rooms, not webinars — proximity to real builders and buyers is a structural advantage.
- Bottom-up individual agent ownership solves the incentive problem that top-down agent mandates cannot.
- Context management — not the harness — is the unsolved problem everyone is building themselves. The company that cracks organizational context wins the next layer.
- Agents are breaking out of engineering in 2026. The companies positioned for non-engineering agents will capture the next wave of enterprise demand.
- Design partner programs beat roadmap speculation. Follow customers with discipline — pluck feathers from every passing goose, but follow no goose absolutely.
- Provocatype-driven development — prototypes built to provoke conversation, not ship features — is the right method when the product shape is genuinely unknown.
FAQ
Why are agent frameworks commoditizing?
The harness layer — prompt routing, tool calling, basic execution — is table stakes that any serious team can build. As model providers ship more capabilities natively and open-source alternatives mature, the framework itself stops being a defensible product. Durable monetization shifts to infrastructure underneath: sandboxing, observability, and orchestration that scale with consumption.
What does "build for the now, not the future" mean for agent companies?
It means capturing the present market rather than skipping it to build for a hypothetical future. Companies that dominate today accumulate users, revenue, and distribution that fund the evolution into whatever comes next. Cursor is a clear example — it captured the IDE moment and is now evolving into agentic coding from a position of strength.
Why does individual agent ownership matter for enterprise adoption?
Top-down agent mandates create a perverse incentive: encoding your knowledge into an agent means training your replacement. Rational employees resist by sandbagging or keeping critical context in their heads. Bottom-up ownership flips this — workers get credit for the agents they build, and their value grows with their agent portfolio.
What is the biggest unsolved problem in agent infrastructure?
Context management. Across dozens of companies building agent systems, the harness is commoditized (mostly Claude Code or open-source forks), triggers and services are bought off the shelf, but context, memory, skills, and orchestration are almost universally built in-house. No vendor has cracked this layer yet, and it is the one piece every company — from Stripe to a five-person startup — is forced to build themselves.
Why are agents about to break out of engineering?
Non-engineering agents are dramatically simpler to build because they do not need environment setup, CI/CD integration, test suites, or stack-specific configuration. They can be ephemeral, spin up and down without complex state, and operate with lightweight tooling. As organizations see what coding agents can do, they will demand the same capabilities for GTM, customer success, operations, and support — and the infrastructure barrier to building those agents is much lower.
What is a design partner program and why does it matter for agent companies?
A design partner program is a structured relationship with early customers who collaborate on product direction through regular syncs, prototype reviews, and shared roadmap influence. For agent companies navigating genuinely unknown product shapes, it prevents two failure modes: building in a vacuum based on internal assumptions, and getting pulled in divergent directions by individual customer requests. The discipline is in synthesizing across partners rather than following any single one.
What is provocatype-driven development?
A provocatype (provocative prototype) is a prototype built not to ship but to provoke conversation about how a tool ought to be used. Instead of building toward a predetermined spec, you create deliberately constrained or unusual interfaces — like "what if the entire product was just a chat box?" — and show them to customers to surface assumptions, preferences, and use cases that traditional requirements gathering misses. It is especially useful when the product shape is genuinely unknown.