The week the vendors caught up

Convergence

On 13 May, Notion shipped its biggest developer surface to date. Workers (hosted code on Notion's plane). External Agents API, with Claude Code, Cursor, Codex, and Decagon as named alpha partners. Agent SDK, so a per-client Custom Agent can be embedded in any HTTPS-callable surface — Notion native, an iPhone Action Button, a web widget, a voice channel. Database Sync, pulling from any API. MCP enhancements 91% more token-efficient. CLI and Markdown API as GA.

The next day, Anthropic shipped Claude for Small Business — first-party connectors to eleven SaaS tools, fifteen pre-built workflows, fifteen first-party Skills across finance, ops, sales, marketing, HR, and customer service. No extra charge above existing subscriptions. A free PayPal-funded fluency course. A ten-city US tour.

Two vendors. Five days. Both arrivals point at the same place — the agent-with-tools shape that Concept 38 filed in late April as the direction enterprise SaaS was already moving. Anthropic's launch extends the thesis one rung deeper: it isn't just enterprise software collapsing under the LLM. The LLM vendor itself is now shipping the packaged convergence to small operators. And Notion's launch closes the other end — Claude Code becomes a named first-class developer surface against the workspace this project's clients live in.

The Skill abstraction Concept 45 rests on stopped being our vocabulary this week. Two market leaders shipped it as theirs.

Compounding

Concept 53 filed on 15 May out of a voice reflection mid-session. The frame: Neuro Support compounds along three multiplicative axes.

Lived experience accumulates linearly with years on the ground. Jamie has fifteen. Wiki and Protocol accumulate with deposits — disciplined ingest, lint, query close — and don't decay if tended. Model capability rises underneath, exogenous, owed nothing back. Three vectors, multiplying. Today's Claude reading today's wiki is one agent. 2031 Claude reading 2031 wiki is a different scale of agent reading a different scale of substrate.

This is the distinguishing property of the service among sole-trader work. Most sole-trader services compound only with hours worked — and hours are bounded by a single life. This one compounds with hours plus accumulated data plus exogenous capability gains. The wiki accrues while Jamie sleeps. Principle 7 protects the model vector — markdown and schema, portable across providers, so the substrate survives a vendor change. The lived-experience vector is the uncopyable moat (Principle 6).

Filed at medium confidence, page-level only per Heuristic 35. No schema propagation until validated by lived practice. The frame is the bet, not the announcement.

Architecture

Concept 51 closed this week into its canonical client-facing names — Neuro Agent (default, unmarked) and Neuro Agent Team (variant). One product, two operating shapes by who's in the Notion teamspace. Internal shorthand 1 and 1+ for the technical layer.

Five install variants drop out of that, depending on which planes are switched on:

• Install only, default — Home plane.
• Install only, Team — Home plus Team.
• Install plus Operational, default — Home plus Neuro Support.
• Install plus Operational, Team active client — Home plus Team plus Neuro Support.
• Install plus Operational, Team passive client — Team plus Neuro Support (Home empty or absent).

The Custom Agent reads whichever planes are present. Output channels render from whichever planes are populated. Same architecture across all five.

Mid-consolidation, Jamie's lived-experience spot-check caught a contradiction. About Me should be on Home, not Support — I thought Concept 27 might be stale. Cross-check against the canon confirmed it. Concept 49's foundation-versus-wrapper split, filed nine days earlier, had retroactively made the old placement wrong: Neuro Support is the wrapper that comes with Operational, Home is the foundation that persists. About Me as client identity must persist regardless of subscription. The wiki had the conflict. The wiki didn't surface it on its own. Lived intuition was the trigger; the cross-check was the proof.

Two products didn't fork. One product clarified.

Mechanism

The session that produced Concept 51 began with a different question. Why are some people genuinely bad at technology, and is it really just exposure? The folk answer (older generations had less access) is incomplete. The empirical answer is older and harder.

Czaja's CREATE study, N=1,204, found fluid intelligence the single strongest predictor of tech adoption — β=0.727, larger than age, larger than education, larger than experience. Lintunen's 2024 study found working memory and executive function effects equal to or larger than experience effects. Computer self-efficacy meta-analyses cluster at r=0.32 to 0.59. Tech competence variance is largely structural, not generational.

Which lands the deeper observation. The same cognitive variables that produce the executive function deficit in ADHD, ASD, intellectual disability, traumatic brain injury, and dementia also produce the tech adoption difficulty. One cognitive bottleneck. Two visible consequences.

The cross-disability convergence Concept 24 named at the positioning layer now has a mechanism. The architectural answer — one product with configurable team access, not two separate products — falls out without fanfare. A TBI survivor with a spouse, a young adult with intellectual disability with a parent, a dementia client with an adult child, an ADHD adult with a supportive partner — all Team variants of the same product. The shape that fragmenting would force has nothing under it to fragment.

Wiki

CLAUDE.md grew from project conception with wiki discipline — append, cross-reference, file the origin paragraph at every layer. That worked beautifully for the wiki. It worked badly for the system prompt, because a system prompt pays per token on every turn forever, while a wiki page pays once when written and again only when read.

A latency diagnosis on Saturday surfaced the cost. Three trim passes followed. CLAUDE.md fell from around 11,700 tokens per turn to around 5,000. Auto-loaded context dropped from 30,000 to 23,000. The relocated detail moved to three new files in Docs/ — naming-conventions for retirement history, kb-operations for the Ingest / Lint / Review workflows, cross-agent-handoffs for the routing model and provenance. CLAUDE.md kept the rules that fire every turn.

Heuristic 40 named the lesson. A system prompt is not a wiki page. Wiki pages tolerate accretion; system prompts don't. Different substrates, different disciplines, same project.

The same week produced a sibling observation worth naming separately. Three regulatory ingests — tax, privacy, lone-worker — ran in one Saturday session and confirmed structural arguments the wiki already held. The HIPC applies. The closely-held company carve-out is exactly what we'd inferred. The lone-working safety pathway lands without contradiction. None of these changed a structural claim. Each slotted underneath as authority. Lateral confirmation — the wiki's external-evidence layer deepening behind arguments that were already correct. A third wiki-value direction beside forward-compounding and backward-gap-surfacing.

And on Sunday, the lint follow-through closed Concepts 11, 28, and 31 of their last retired-terminology debt — Concept 43's canon propagated cleanly across the pricing surface. A queue that had been carried for weeks finally cleared.

What's next

The Skill abstraction got two-vendor validation. Concept 51 closed. Concept 53 named the compounding. The wiki audited its own drift in the same session it compounded.

Next: pin Concept 43's Max billing rhythm. Touch the foundational pages flagged by Sunday's staleness audit when reasons land. Watch Notion's post-beta credit pricing, due 11 August. Issue 09 lands on 24 May.