Make Something Agents Want: 3 Conditions for Software to Survive in the Agent Era

YC's philosophy has shifted from "Make something people want" to **"Make something agents want"**. In an era where agents choose, recommend, and deploy, the value axis of software has fundamentally moved. YC Lightcone,...

Table of Contents

Summary

YC’s philosophy has shifted from “Make something people want” to “Make something agents want”. In an era where agents choose, recommend, and deploy, the value axis of software has fundamentally moved. YC Lightcone, a16z, and Lablup’s Shin Jeong-gyu — three sources all pointing in the same direction — converge on 3 conditions for surviving the agent era: (1) documentation that agents choose, (2) a harness rather than code, (3) an unreplicable domain.

The core message is value migration: code → documentation → harness → domain. The value of code itself converges to zero, and the true moat lies in domain knowledge and customer lock-in.

Key Concepts

1. Agent-Chosen Documentation

  • What: When an agent (LLM) selects a tool or service, documentation quality is the decisive criterion
  • Why: Agents understand the world through APIs and documentation, not landing pages
  • Impact: A 5% difference in documentation quality → several times the difference in customer count
WinnerLoserDifference
SupabaseCompeting DBs”Best documentation” — mentioned by YC Lightcone
ResendSendGrid (10,000+ employees)Built LLM.txt, agent-friendly documentation → ChatGPT inbound Top 3
MintlifyA tool that converts all developers’ documentation into agent-friendly format
  • New formula: SEO era = Google brings customers → Agent era = documentation brings customers
  • Actionable insight: LLM.txt, structured API documentation, code-snippet-first documentation design

2. Harness, Not Code

  • What: A context + workflow system that makes agents work automatically
  • Why: Code is churned out by agents and models get swapped out, but harness accumulates organizational know-how
  • Impact: “The value of code converges to zero, but the harness that produces it is the new definition of software” — Shin Jeong-gyu

Lablup case (40 days, 1 million lines):

ComponentRole
CLAUDE.mdContext definition (SOUL Document)
PROGRESS.mdProgress tracking
PLAN.mdPlan management
Cron (every 15 min)Issue analysis → code generation → PR → merge automation
Sub-agents (up to 50)Parallel processing, humans review only

Non-technical roles case:

RoleOutcome
CFO30 minutes of learning → 2-hour task reduced to 3 minutes
Content manager250 documents converted in 1 week + news crawling automation harness built
  • Key distinction: They didn’t learn to code — they built a harness

3. Unreplicable Domain

  • What: Documentation and harness will eventually be caught up to → the final moat is domain knowledge
  • Why: 10 years of edge cases + customer lock-in that general-purpose AI cannot replicate all at once
  • Impact: a16z data — Vertical SaaS beats Finance/ERP/Marketing/Productivity across the board

“Install the waterwheel where the water falls hardest” — Shin Jeong-gyu

DomainWhy It Cannot Be Replicated
Construction settlement logicComplex subcontracting structure, conventions, exception handling
Medical insurance claimsRegulations, code systems, review criteria
Logistics dispatch optimizationReal-time variables, driver patterns, regional characteristics

Replicable vs. Unreplicable:

Replicable (caught up quickly)Unreplicable (true moat)
Tool proficiencyKnowledge of customer failure patterns
Prompting techniques (spreads within a week)Memory of problems solved together over years
Agent-friendly documentation (Mintlify arrives)Knowing what that organization will never be able to do
Harness patternsEdge cases accumulated over 10 years

Practical Applications

Use Case 1: Agent-Friendly Documentation Design

  • Add LLM.txt to API documentation (reference the Resend case)
  • Place code snippets at the top of documentation
  • Design structured metadata for easy agent parsing
  • Evaluate adopting tools like Mintlify

Use Case 2: Building an Organizational Harness

  • Introduce a SOUL Document system based on CLAUDE.md / PROGRESS.md / PLAN.md
  • Build a GitHub Issue → PR → Merge automation pipeline
  • Train non-technical roles to build harnesses (not coding education — harness education)

Use Case 3: Evaluating Domain Moat

  • Measure the “IT-to-domain gap” of your own service
  • Catalog unreplicable domain assets (edge cases, customer relationships, regulatory knowledge)
  • Explore Vertical SaaS opportunities: construction, healthcare, logistics, and other high-gap areas

Value Migration Framework

Code
  ↓ Agents can generate it → value converges to zero
Documentation
  ↓ Democratized by tools like Mintlify → caught up to
Harness
  ↓ Once patterns spread → caught up to
Domain Knowledge
  → 10 years of accumulation + customer lock-in = true moat

Survival conditions checklist:

  1. Do you have documentation that agents choose?
  2. Are you accumulating a harness rather than just code?
  3. Do you have an unreplicable domain moat?

Limitations & Gotchas

  • Documentation quality alone cannot serve as a long-term moat (tools like Mintlify level the playing field)
  • Harness patterns will eventually spread — the differentiation window is limited
  • The “agent era” does not apply equally to all B2B SaaS — the impact is felt first in developer tooling where agent penetration is highest
  • Even Vertical SaaS moats are not permanent — domain-specific LLMs may emerge and narrow the gap

References

Next Steps

  • Research LLM.txt spec and evaluate applying it to current projects
  • Experiment with the SOUL Document system (CLAUDE.md / PROGRESS.md / PLAN.md)
  • Evaluate current project’s agent-friendliness (documentation structure, API design)
  • Research Vertical SaaS opportunity areas (based on a16z data)
  • Evaluate Mintlify adoption or establish internal agent-friendly documentation guidelines

Notes: The core of this post is “value migration.” The framework that value moves along the path of code → documentation → harness → domain is an extremely practical strategic guide for anyone currently building AI agent tools. In particular, the harness concept connects directly to Claude Code’s CLAUDE.md-based workflow, and is useful for distinguishing what is already being practiced from what needs to be strengthened.