Strategy 4: Parallel Entry — Foundation Models

The five technologies in Strategy 3 climbed the stack one layer at a time, each faster than the one before. Foundation models do not climb — they enter all seven layers at once, because each of those layers is now implemented in software that a foundation model can read, reason over, and act on. The cognitive worker is the first primitive in the stack's history that occupies multiple layers simultaneously, replicates at zero marginal cost, and arrives before the institutional substrate that should constrain it.

Part 1 — The Primitive

Three properties define the cognitive worker in 2026. Each is grounded in shipping products and measured trajectories. Each compounds when combined with the others.

Capability

88.6% → 95.0%

SWE-bench Verified, Claude Opus 4.8 → Fable 5, on real GitHub issues

METR's time-horizon doubles every 4.3 months. In 2024 agents reliably handled tasks taking humans minutes; by mid-2027, multi-day tasks become tractable. Note: 17–21 pt gap exists between vendor and standardized scaffolds — anchor low for deployment, high for trajectory.

Cost

$20 → $0.07

Per million tokens at GPT-3.5-MMLU equivalent quality, Nov 2022 → Oct 2024 (Stanford AI Index)

Inference price halves every two months (Epoch AI, median 50× annual decline). The economic argument that depends on cognitive workers staying expensive has a two-year shelf life. The argument that depends on them being cheap is, by 2028, an argument about labor whose unit-economic floor approaches zero.

Protocol

MCP + A2A

97M monthly SDK downloads, 150+ partner orgs — adopted in 12 months, donated to Linux Foundation

TCP/IP for cognitive workers. MCP standardizes the model-to-tool boundary; A2A standardizes the agent-to-agent boundary across firms. Without these, cross-vendor and cross-firm cognitive work could not happen. With them, it is technologically permitted — though not yet ubiquitous.

Part 2 — Parallel Entry Across the Seven Layers

For each layer: what is shipping in mid-2026, how the layer is being redefined (not just traversed), and where the curatorial responsibility sits for the human who deploys the worker.

Task Acceleration

Shipping in 2026

GitHub Copilot 20M users · 90% F100 Cursor $2B ARR Claude Code $1B in 6mo ChatGPT 1B MAU M365 Copilot 160M ent. users Harvey $11B val OpenEvidence 40% of US physicians

+14% productivity avg (Brynjolfsson) vs −19% for experts on familiar code (METR). Heterogeneity is real.

How the layer is redefined

Prior tech kept the worker; the task sped up. Cognitive workers convert the bounded individual task into something deployed at parallel scale. The same prompt that drafts one memo drafts ten thousand. Individual judgment is now encoded into a process that runs against a population.

Curatorial responsibility

What to delegate. The simplest decision in the stack — and the one where bad judgment compounds fastest, because the same prompt drafts ten thousand bad memos as easily as ten thousand good ones.

Delegated Execution

Shipping in 2026

Computer Use Oct 2024 Operator Jan 2025 Devin 13.86% → 95% SWE-bench Replit Agent 3 200min autonomy Sierra $15.8B val · $150M ARR Decagon $4.5B val Intercom Fin 3 67% resolution

Klarna's 700-agent equivalent + 2025 partial reversion. Substitution at the customer-trust boundary is contestable, not monotonic.

How the layer is redefined

Prior tech delegated procedure — the thermostat, the cron daemon. Cognitive workers delegate judgment. A Sierra agent decides whether to refund, escalate, or clarify when listening to a customer describe a problem the system has never seen. The deploying human cannot, even in principle, audit every decision in advance.

Curatorial responsibility

Where to draw boundaries. What authority is this worker permitted, against what evidence, with what escalation? Klarna's case study: the agent had authority to resolve; what it lacked was the boundary that escalated when frustration exceeded its competence.

Persistent Environments

Shipping in 2026

AWS Bedrock 180% YoY Azure AI Foundry Vertex AI Apple Intelligence iOS 18.1 Copilot+ PCs ChatGPT Memory 41.5%→67.9% recall Letta agent OS Llama 4 DeepSeek R1 27× cheaper

Open vs closed gap shrank from 8% → 1.7% in 12 months (Stanford AI Index).

How the layer is redefined

Prior tech: persistent environments were infrastructure the worker stood on. Cognitive workers are themselves the infrastructure. The tool is the platform. When inference halves every two months, an always-on agent runtime becomes baseline — not a research project.

Curatorial responsibility

Architecture and governance. What runs where, what persists across sessions, what is local versus cloud, which model serves which class of decision. The closed-vs-open choice is no longer just cost or capability — it is who has access to the worker's memory.

Organizational Coordination

Shipping in 2026

Slack AI · Teams Copilot · Notion AI Granola $1.5B val Glean $200M ARR · $7.2B val Goldman GS AI 46.5K employees Anthropic on Anthropic 80% code authored by Claude Shopify mandate AI in performance reviews LangGraph · CrewAI

A2A: 150+ partner orgs across SAP, Salesforce, ServiceNow, Workday, Intuit, Box.

How the layer is redefined

Prior tech: humans coordinated through artifacts (memo, email, Slack). Cognitive workers coordinate with humans — increasingly as peers, not tools. The Anthropic engineer spending 59% of work hours operating through Claude is collaborating with a teammate, not using a tool.

Curatorial responsibility

Role design and authority allocation. Which decisions does the worker make alone? Which require human concurrence? Where does its office end and the human manager's begin? The Layer 4 curator is the org designer.

Market & Social Intermediation

Shipping in 2026

Amazon Rufus 300M users · ~$12B incr. sales ChatGPT Shopping + Instant Checkout Sep 2025 Perplexity $450M ARR · $20B val Google AI Overviews 60% of queries · 83% zero-click LinkedIn Hiring Assistant Meta Advantage+ $60B+ run-rate

Despite the hype, Google still held 90.39% of global search share (May 2026). Agents added to the market more than they displaced.

How the layer is redefined

Prior tech compressed buyer-seller distance — catalog, search, listing. Cognitive workers make the decision. The Rufus user buying at 60% higher completion isn't finding a product faster; they are buying what the agent recommended. The conflict of interest is structural — the worker is paid by one party, advising another.

Curatorial responsibility

Intent translation. Whose interest wins when the consumer's stated request and the consumer's underlying intent diverge? This is where the moral seriousness of deployment is highest.

Economic Settlement

Shipping in 2026

ACP Sep 2025 · OpenAI + Stripe AP2 Sep 2025 · 100+ partners x402 119M tx on Base Mastercard Agent Pay Visa Intelligent Commerce PayPal Agentic Stripe Bridge $20B+ stablecoin Okta + Auth0 for Agents

None of these existed in 2023. All GA or near-GA by mid-2026. The rails have outpaced every prior settlement build.

How the layer is redefined

Prior tech: settlement asked which artifact records the obligation. Cognitive workers add: which artifact represents the agent's authority to commit on the human's behalf. The agent is now a settlement principal — holds a wallet, signs an authorization, settles an obligation.

Curatorial responsibility

Authorization design. AP2's Intent / Cart / Payment Mandate framework is the first systematic attempt to make this curation explicit — the rules under which an agent may spend, the limits requiring escalation, the cryptographic boundary defining what the human is committing to.

Shared Epistemic Infrastructure

Shipping in 2026

C2PA Content Credentials SynthID watermarking EU AI Act €35M / 7% turnover Community Notes at Meta Jan 2025 DoNotPay FTC precedent US + UK AISI RSP · Preparedness Framework

AI-generated articles climbed from 2.2% (Jan 2020) to 51.7% (May 2025) of Common Crawl URLs. Trust institutions are forming, but lagging.

How the layer is redefined

Prior tech: shared epistemic infrastructure grew slowly around durable institutions — canon, peer review, accredited fact-checker. Cognitive workers mediate both production and consumption of authoritative knowledge. The paper, the news article, the search result are all increasingly the worker's synthesis. The institutions to ask who generated what, with what evidence, under whose instruction are not yet built.

Curatorial responsibility

What the worker treats — and teaches — as authoritative. Training data, citation behavior, the willingness to refuse uncertain claims, transparency about its own limits. The deepest curation in the stack.

Part 3 — The Settlement–Trust Inversion

The single structural feature that makes this transition different from all five prior ones. Settlement rails (L6) are forming at unprecedented speed. Trust rails (L7) are barely formed. Every prior technology resolved this the other way around.

Layer 6 — Settlement Rails

FORMING FAST

~88% mature

ACP, AP2, x402 — all GA or near-GA, none existed in 2023
Mastercard Agent Pay, Visa Intelligent Commerce, PayPal Agentic
Stripe Bridge — $20B+ annualized stablecoin volume
x402 on Base — 119M+ transactions, $600M annualized
Okta & Auth0 for Agents — enterprise IAM
MCP & A2A — Linux Foundation governance

Layer 7 — Trust Rails

BARELY FORMED

~22% mature

No cross-vendor agent reputation system at scale
No AI-liability insurance product at scale
No SOC-2-for-agents attestation standard
No portable cross-vendor identity for cognitive workers
No settled answer to "if an agent harms you, whom do you sue"
No cognitive notary function — who attests which agent did what

Writing took millennia at Layer 7 because the institutions had to form. Electricity, the computer, and search inherited those institutions and still spent decades. Cognitive workers are the first technology in stack history to settle before they have trust. The L6 rails are running ahead of the L7 institutions that should constrain them. This is the structural fragility of the current transition — and the locus where institutional work has the most leverage.

Part 5 — Curatorial Responsibility, Layer by Layer

The historical lesson from `comparison_5_stack.md`: capability becomes cheap, judgment compounds. Here is the curation that compounds at each layer.

Task Acceleration

What to delegate.

The simplest decision in the stack — and where bad judgment compounds fastest. METR's −19% vs Brynjolfsson's +34% says: it matters which tasks.

Delegated Execution

Where to draw boundaries.

What authority is the worker permitted? Against what evidence? With what escalation path? Klarna's reversal was a boundary-drawing failure, not a capability one.

Persistent Environments

Architecture and governance.

What persists where, under whose control? Closed vs open is no longer cost — it is who can audit the worker's memory.

Organizational Coordination

Role design and authority allocation.

Which decisions does the worker make alone? Which require human concurrence? The Layer 4 curator is the org designer.

Market Intermediation

Whose intent the worker translates.

When the consumer's stated request and underlying intent diverge, whose interest wins? The structural conflict of paid-by-one, advising-another.

Economic Settlement

Authorization design.

The cryptographic boundary that defines what the human is committing to. AP2 Mandates and ACP tokens are first systematic attempts.

Shared Epistemic Infrastructure

What the worker treats — and teaches — as authoritative.

Training data, citation, refusal, transparency. The deepest curation in the stack. Shapes what an entire society treats as credible.

Reader Journey — Seven Sections

§1

The Primitive

Define the cognitive worker concretely: capability (88.6%→95% SWE-bench, 4.3mo doubling), cost ($20→$0.07/M tokens), protocol (MCP + A2A). Three properties that compound.

"This isn't a chatbot. It's something else."

§2

Parallel Entry

Walk the seven layers. For each: what's shipping in 2026, how the layer is redefined, where curatorial responsibility sits. Evidence-dense.

"All seven, at once. Not climbing — populating."

§3

The Inversion

The single structural feature that makes this transition different. Settlement rails form first; trust rails barely form. The inverse of every prior technology.

"Wait — this is structurally unprecedented."

§4

Against the Five

5,000 → 4,000 → 120 → 80 → 30 → parallel. The cognitive worker is not the next point on the curve. It is off the curve.

"The pattern broke. Now what?"

§5

What Compounds

Capability is becoming cheap. Judgment about delegation, boundaries, authorization, attestation — that compounds. The work that survives the next model release.

"Judgment is the leverage, not capability."

§6

Curatorial Map

Seven layers, seven curatorial questions. What to delegate, where to draw boundaries, architecture, role design, intent translation, authorization, what to teach as authoritative.

"I see where my work actually lives."

§7

The Window

L1 is crowded. L4–L7 are uncrowded, hard, and compound across the entire transition. The institutions that form around cognitive workers will be shaped by the builders who engage with the formation.

"The window is open and short."