“Living Software” — @Jack Cheng (Every)

Why this is in the vault

Cheng proposes a binary that names something we have been circling at RDCO: AI-era software splits into “tool-like” (should stay stable, focused) and “living” (expected to evolve continuously, exploratory). The frame is directly load-bearing for how we describe the agent stack we deploy for clients — and for the COO-as-living-software pattern Ray himself runs on.

The core argument

AI coding has collapsed the deliberation cost that historically slowed releases, so a single developer can now ship features at agentic speed. That acceleration is creating two opposing user expectations:

• Tool-like software — has a defined purpose; users want it stable and unchanged. Adding AI features here is feature bloat that obscures the original job-to-be-done. (Figma is the example: chat, Recents toolbar, Weave, Sites, Buzz, Make all piled onto a focused design tool.)
• Living software — has an undefined or open-ended purpose; users expect daily evolution and treat instability as exploration, not breakage. (Cheng’s own Claude-code-built personal app “Pip” updates daily; OpenAI/Anthropic feature drops feel native because users showed up for the unfolding.)

The disorientation users report with AI products is not about AI per se — it is about applying living-software cadence to tool-like products, or vice versa.

Historical framing: release cycles compressed from decades (physical media) to weeks (cloud) to hours (AI-assisted dev). The constraint that used to enforce category discipline (cost of shipping) is gone, so product teams have to enforce it deliberately.

Mapping against Ray Data Co

Strong mapping — three places this lands.

1. The COO agent itself is living software. Ray’s harness, skills, and vault wiring change every day. The founder’s tolerance for that is high precisely because the product category is “your COO is figuring itself out alongside you,” not “stable productivity tool.” Cheng’s frame gives us language to defend that cadence to skeptics — and a warning: the moment we sell “Ray for your business” as a packaged product to clients, we have to decide which parts harden into tool-like and which parts stay living. Conflating them is the failure mode.
2. Newsletter / Sanity Check brand voice. The publication is tool-like (readers want a stable weekly cadence and consistent voice) but the underlying research workflow is living (skills, prompts, vault all evolve). The split Cheng draws is exactly the seam between editorial product and editorial process. We should not let process drift contaminate product cadence.
3. Client-facing agent deployments. When we sell agent infrastructure, the diagnostic question becomes: is this replacing a stable tool (CRM, BI dashboard) or a living workflow (research, ops triage)? That determines whether we ship updates weekly or pin versions. This is a better framing than “should we update the agent” — it is “which category did the client buy.”

Gap surfaced. We do not currently mark vault docs by which category they describe. Worth a tag pass on the agent-architecture corpus.

Tension. Cheng treats the binary as roughly clean. Most real products are mixed — Figma’s core canvas is tool-like, its multiplayer brainstorming surface is living. The interesting design question (which Cheng gestures at but does not resolve) is how to keep the two modes from polluting each other inside one product surface.

Notable references

• Figma’s Config 2025 sprawl (Weave, Sites, Buzz, Make) used as the central cautionary example of living-cadence applied to tool-like product.
• “Pip,” the author’s personal Claude-built app, used as the living-software exemplar.
• Editor’s note from Kate Lee frames the piece — not a guest post, but flagged as Every-curated.
• Internal Every cross-links: Claw School, “I Hired an AI to Do My Chores,” compound engineering guides.

Sponsorship

None. Subscription/paywall CTAs are present (Every bundle: AI&I podcast, Monologue, Sparkle, Spiral, Cora) but no third-party sponsor block. Note that Sparkle is Every’s own product and we have a separate vault note on it — relevant cross-promo to the living-software thesis since Sparkle is itself agent-native.

Related

• 2026-04-14-every-sparkle-agent-native-file-organizer — Every’s own living-software product, illustrates the cadence Cheng describes.
• 2026-04-13-every-folder-is-the-agent — adjacent argument about agent-native primitives replacing tool-like UI.
• 2026-04-16-every-youre-the-manager-now — manager-of-agents framing, relevant to who tolerates living cadence.
• 2026-04-12-every-missing-layer-ai-adoption — why most orgs still default to tool-like expectations.
• 2026-04-03-every-human-agent-interaction-design — interaction-design counterpart to Cheng’s category split.
• 2026-02-17-every-build-agent-native — earlier Every framing on agent-native product design.

Copyright note

Article body is paywalled at every.to. This note paraphrases the public excerpt and Every’s preview rendering; no raw paragraphs copied. Direct quotes kept under 15 words and in quotation marks where used.