Long Live Hard SaaS

Tokens can be commoditized, but product time cannot. AI can make software cheap to draft; it does not make real usage, integration coverage, maintenance, and bug discovery instantaneous.

The thesis

Kevin's launch-period Dedalus Machines observation is that the small bugs agents are finding are evidence for SaaS, not against it. You can prompt an idea into existence, but you cannot shortcut the time required to live inside a system and find the bugs hidden between integration boundaries, user flows, state transitions, billing edges, auth paths, previews, filesystem persistence, and deployment environments. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

The strong version: if AGI could one-shot arbitrary software without bugs, SaaS would be dead. That would require something close to globally proving behavioral correctness for messy, real-world systems. Kevin frames that as halting-problem-class difficulty: not a claim that every product bug literally reduces to the halting problem, but that general "bug-free arbitrary software" is not just a generation task. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

Until then, hard SaaS companies survive because they sell a maintained system. The value is not only the code artifact. It is dedicated software that solves a real problem, is constantly maintained by others, and removes overhead and friction for the user. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

Why time remains scarce

AI compresses drafting, scaffolding, and local reasoning. It does not compress the full calendar time of discovering how a product behaves under real usage. Some bugs only appear after:

• a user repeats a flow enough times to hit an odd state
• auth, billing, storage, preview URLs, and external APIs interact
• migrations meet existing data
• long-running agents accumulate filesystem state and partial work
• support, docs, onboarding, and product copy reveal mismatched assumptions
• a fix for one issue creates a regression somewhere else

This is why What Is Engineering matters. Research can isolate a question; engineering keeps the product working as a living system. SaaS companies are paid to own that evolving maintenance burden so users do not have to. Source: What Is Engineering; raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

Mythos and the bug-finding limit

The Mythos/Glasswing discussion sharpens rather than weakens the thesis. Cloudflare's Project Glasswing writeup says Mythos improved vulnerability work by chaining primitives and generating proofs, but the same post also says generic coding agents are the wrong shape for broad vulnerability coverage. Cloudflare moved to a harness: narrow scopes, adversarial validation, split reasoning chains, parallel tasks, dedupe, tracing, and structured reports. Source: Cloudflare Blog, 2026-05-18

That is not "the model magically finds all bugs." It is Harness Engineering with a very expensive reasoning engine inside it. The Slack discussion's distinction is important: security bugs and SaaS-development bugs overlap, but end-to-end product bugs are often integration failures, lifecycle failures, and user-experience failures, not isolated logical vulnerabilities. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md; Cloudflare Blog, 2026-05-18

Cloudflare also reports that letting the model patch findings produced fixes that could break other dependencies. That is the hard SaaS lesson in miniature: finding and fixing are not enough; regression, rollout, and architecture still matter. Source: Cloudflare Blog, 2026-05-18

Councils, loops, and token burn

LLM Council and OpenRouter Fusion are attempts to wrap model weaknesses with orchestration: run panels, delegate across models, judge outputs, and loop with verification. OpenRouter's own Fusion FAQ says its benchmark does not cover long-horizon tasks and that Fusion is not a drop-in coding-model replacement. It is a selective tool for questions that merit multiple perspectives. Source: OpenRouter Blog FAQ, 2026-06-14

That supports Kevin's point. Long-reasoning systems can buy more persistence by spending more tokens and wall-clock time. They do not erase the need for a product to be used, observed, patched, and re-tested across the nooks and crannies. In practice, the best systems are loops plus harnesses plus product time, not bare models. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

Dedalus Machines implication

Dedalus Machines is a clean example because the product is not a static code blob. It is a persistent machine abstraction with microVM isolation, filesystem state, sleep/wake semantics, billing, command execution, preview URLs, SDK behavior, docs, and launch onboarding. Bugs found during launch prep are the product revealing its real shape. Source: Dedalus Machines - Positioning & Pitch; raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

The launch lesson: agents should expand the QA surface, not create the illusion that QA is solved. A good agent fleet can crawl more flows, create more repros, inspect more code paths, and leave better traces. The company still wins by maintaining the system over time and turning every found bug into a guardrail, test, doc, dashboard, or product simplification. Source: Dedalus Labs; Empirical Verification; raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

Essay seed

Working title: Long Live Hard SaaS.

Core argument:

1. AI commoditizes token generation, not time-in-system.
2. The real moat of SaaS is maintained correctness under real usage.
3. Mythos-style bug finding proves harnesses matter; it does not prove one-shot software is solved.
4. LLM councils and long-reasoning agents are compute-heavy workarounds for bare-model limits.
5. The winner is the company that turns use into fewer bugs for everyone else.

Potential hook: "My agents are finding bugs in Dedalus Machines before launch. That makes me more bullish on SaaS, not less."

Potential closer: "Until an AGI can one-shot arbitrary bug-free software, long live hard SaaS companies."

---

Timeline

• 2026-06-22 | Kevin articulated the thesis during Dedalus Machines launch prep: tokens may be commoditized, but time-in-system cannot be skipped; hard SaaS survives because bug discovery, integration coverage, and maintenance are durable work. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md