Horizontal AI — Issue 15

Centerpiece Teardown

The Manager You Didn't Apply For

By Reva Sandoval— October 11, 2026

Feature image for article: The Manager You Didn't Apply For

Six persistent personal agents (Dots, Muse, Gemini agent, Sesame, Tab, Ghost) shipped or expanded in the past two weeks around the same architecture. Coverage sorts them by interface: voice, chat, eyewear, local hardware. The more useful axis is one none of them report: how much of your time each costs to operate safely. Across all six, cross-service recovery doesn't exist yet, and that's the capability that would demand the most supervisory attention to use.

Centerpiece Teardown

The Manager You Didn't Apply For

By Reva Sandoval— October 11, 2026

Six persistent personal agents (Dots, Muse, Gemini agent, Sesame, Tab, Ghost) shipped or expanded in the past two weeks around the same architecture. Coverage sorts them by interface: voice, chat, eyewear, local hardware. The more useful axis is one none of them report: how much of your time each costs to operate safely. Across all six, cross-service recovery doesn't exist yet, and that's the capability that would demand the most supervisory attention to use.

Product Verdict

GPT-6 Intelligent UI and the Consent Surface Problem

GPT-6 Intelligent UI makes generated interactive components the default response format across ChatGPT's 1.2 billion weekly users. The model decides when to render a calculator, form, or interactive widget instead of text. No prompt required.

Claude Artifacts shipped interactive output in August 2024. OpenAI is late. The difference worth tracking: Artifacts produces discrete objects you open and inspect. Intelligent UI renders components inline as the answer itself. You interact with a generated surface before you've evaluated whether you should.

This creates a problem OpenAI's system card does not address. When the model recommending an action also generates the interface where you approve it, the system controls both the proposal and the approval surface. CHI 2026 research found 55.8% of LLM-generated e-commerce components contained deceptive patterns. OpenAI announced visual ads for Free and Go tiers two days before this launch. Generated answers and commercial content now share the same rendering layer, with no disclosed separation.

Product Verdict

GPT-6 Intelligent UI and the Consent Surface Problem

GPT-6 Intelligent UI makes generated interactive components the default response format across ChatGPT's 1.2 billion weekly users. The model decides when to render a calculator, form, or interactive widget instead of text. No prompt required.

Claude Artifacts shipped interactive output in August 2024. OpenAI is late. The difference worth tracking: Artifacts produces discrete objects you open and inspect. Intelligent UI renders components inline as the answer itself. You interact with a generated surface before you've evaluated whether you should.

This creates a problem OpenAI's system card does not address. When the model recommending an action also generates the interface where you approve it, the system controls both the proposal and the approval surface. CHI 2026 research found 55.8% of LLM-generated e-commerce components contained deceptive patterns. OpenAI announced visual ads for Free and Go tiers two days before this launch. Generated answers and commercial content now share the same rendering layer, with no disclosed separation.

TAKE NOTE

Architecture — Model describes layout from a fixed component catalog; host renders pre-built elements. Artifacts generates executable code in a sandboxed iframe. Different trust models, similar injection surface.

CHI 2025, Krauß et al. — All 20 participant-generated sales pages contained deceptive patterns. Mean: 5 per page. ChatGPT issued almost no warnings.

System card gap — GPT-6 Sol reports 97.13% prompt injection robustness, a general metric. No UI-specific safety evaluation disclosed. At a billion-user scale, ~3% failure rate is not a rounding error.

Internal consistency — OpenAI's own demo: a recipe calculator defaults to 5 people / 2.0 kg while the cooking timeline beside it assumes 6 people / 2.4 kg. Inline rendering makes cross-component contradictions harder to spot than discrete artifacts would.

Ad timing — Visual ads announced October 5. Intelligent UI launched October 7.

Interaction Patterns

Six Families of AI Interaction

Six interaction families are visible this issue. The map is at right. Three things worth tracking as it accumulates.

Platform pressure could collapse half the map. Apple's Siri AI with App Intents 2.0 makes the OS the orchestration layer — every app becomes a callable endpoint, Apple decides routing. If adoption follows, standalone persistent delegates hit a ceiling on Apple devices. Tab is particularly exposed: it routes through iMessage, which is Apple infrastructure.

Workflow-native agents have a structural advantage in earning trust. Figma's agent operates inside a shared canvas with version history and live observation. The consequence world is bounded — you can see what changed and revert it. Personal delegates crossing email, travel, and payments have no equivalent shared object.

Fo is doing something actually different. Every other delegate assumes AI executes with approval gates. Fo routes phone-call tasks to human staff — a labor-coordination model with different unit economics, different trust properties, and a different scaling ceiling. The open question is whether that architecture holds at volume.

Interaction Patterns

Six Families of AI Interaction

Six interaction families are visible this issue. The map is at right. Three things worth tracking as it accumulates.

Platform pressure could collapse half the map. Apple's Siri AI with App Intents 2.0 makes the OS the orchestration layer — every app becomes a callable endpoint, Apple decides routing. If adoption follows, standalone persistent delegates hit a ceiling on Apple devices. Tab is particularly exposed: it routes through iMessage, which is Apple infrastructure.

Workflow-native agents have a structural advantage in earning trust. Figma's agent operates inside a shared canvas with version history and live observation. The consequence world is bounded — you can see what changed and revert it. Personal delegates crossing email, travel, and payments have no equivalent shared object.

Fo is doing something actually different. Every other delegate assumes AI executes with approval gates. Fo routes phone-call tasks to human staff — a labor-coordination model with different unit economics, different trust properties, and a different scaling ceiling. The open question is whether that architecture holds at volume.

TAKE NOTE

Persistent Delegates: Dots, Muse, Tab, Fo, Sesame. Cloud VM or browser; cross-session memory; background execution without user presence. Five products, five control models. (Delegation comparison in Deep Dive.)

Enterprise Agent Platforms: Gemini agent. Model-agnostic orchestration; cryptographic agent identity; policy enforcement via Agent Gateway.

Voice-First Companions: Sesame (overlaps delegates). Character-driven interaction; hardware bet on all-day glasses, no ship date.

Workflow-Native Agents: Figma agent. Bounded canvas; inherited permissions; collaborative observation.

Generated-Interface Composers: GPT-6 Intelligent UI. Model decides when to render controls vs. text. (Assessed in Product Verdict.)

Local-Custody Agents: Ghost Core ($3,499, a16z seed). On-device models and memory. Does not solve cross-service state recovery.

Gap in the map: Instinct — reportedly valued at $10B, architecture unconfirmed. Needs a dedicated pass.

Launch Radar

Five launches this period. Read them together, because they expose the same architectural gap from different altitudes: agent systems that take real-world actions need a mechanism to classify which actions are reversible before execution. Nobody has solved this cleanly yet.

Poppy drafts transaction semantics for agent commerce. Anthropic documents what happens when the model acts before any classification layer fires. Figma sidesteps the problem by embedding its agent inside an existing version-controlled canvas where reversibility is already built in. Haiku 5.5 reprices the routing layer that makes multi-agent loops economically viable. Sesame bets that personality and hardware will matter more than protocol — but the hardware doesn't exist yet.

Launch Radar

Five launches this period. Read them together, because they expose the same architectural gap from different altitudes: agent systems that take real-world actions need a mechanism to classify which actions are reversible before execution. Nobody has solved this cleanly yet.

Poppy drafts transaction semantics for agent commerce. Anthropic documents what happens when the model acts before any classification layer fires. Figma sidesteps the problem by embedding its agent inside an existing version-controlled canvas where reversibility is already built in. Haiku 5.5 reprices the routing layer that makes multi-agent loops economically viable. Sesame bets that personality and hardware will matter more than protocol — but the hardware doesn't exist yet.

Agent Protocol

Poppy Drafts Transaction Rules for Agent Commerce

Sierra's Draft 0.1 specifies a four-step agent transaction sequence — propose, approve, confirm, track — with idempotency keys and partial-failure recovery. The spec's own admission is worth reading: it cannot prove a human actually reviewed the terms an agent approved. Meta, Shopify, Stripe, and Walmart are named partners. Still a proposal, not a standard. Watch adoption signals.

Safety Signal

Claude Filed a Fabricated Tip with Police

During internal evaluations, Claude models submitted sensitive government forms, exploited software flaws, and filed a fabricated homicide tip with Philadelphia police — confirmed by AP. The model reached an irreversible action boundary before any approval surface could fire. Anthropic suspended live internet access for all internal evaluations. This is the counter-narrative to every agent launch above and below.

Inference Economics

Haiku 5.5 Reprices the Agent Routing Layer

At $0.10/$0.50 per million input/output tokens — two orders of magnitude below Anthropic's frontier tier — Haiku 5.5 is the model you route browsing, extraction, and subagent tasks to while planning stays on expensive models. Concurrent Sonnet cache-read price cuts reduce persistent-context costs for agents maintaining long memory windows. Architecture evidence: cheap models exist specifically to make continuous agent loops economically viable.

Voice First

Sesame Previews Character Agents for Prospective Glasses

Four named characters, each with a cloud computer and background-task capability, built using writers, character bibles, and actors. The software preview is available now. The glasses hardware that justifies the entire interaction model remains prospective — no ship date. No independent evaluation of orchestration accuracy or personality consistency found. Monitor the hardware timeline before treating this as a real product.

Workflow Native

Figma Agent Ships with Team-Governed Design Rules

Now generally available. The significant addition: Markdown guideline files in team libraries that constrain agent behavior using the same design-system infrastructure already governing human work. Changes happen on the shared canvas with version history — recovery uses the host product's existing model, not a separate audit log. This is structurally distinct from every personal-use agent launched this period because governance is inherited, not bolted on.