Destination vs. Backend: The Research-Tool Market Bifurcates — and Both Layers Converge on MCP
Revised 2026-09-23: adds OpenAI's ChatGPT Work as the newest host-layer compression move on the destination side, and links the Amazon–Muse blockade as the gated-source squeeze (Agentic Browsing and the CFAA: The Ninth Circuit's Amazon v. Perplexity Ruling — and the Post-Legal Blockade of Agents).
The open question this cycle set out to resolve: if sources become MCP backends and models become commodities, does durable value sit in the persistent workspace/orchestrator layer (Hey, Lefty's bet) or in the destination and backend layers? Two developments this week sharpen the squeeze on the middle:
1. The host layer keeps compressing the destination layer from above. OpenAI now ships ChatGPT Work — per its own help documentation, "Work is an agent designed for longer, multi-step work and finished deliverables," with the Work use case enumerated as "Research a topic, analyze information, or create a document, spreadsheet, presentation, report, or Site." It is bundled into existing subscriptions (Plus and up), runs cloud or local, inherits Projects context ("A project keeps related chats, files, and instructions together"), and can "run once, repeat on a schedule or trigger, or monitor for changes through Scheduled Tasks" — including webhook triggers on Gmail, Slack, and GitHub events1. This is still chat-adjacent persistence — no notebooks, no citation library, no agent-built views — so the standing host-workspace watch has not fired. But the destination-layer wedge for standalone research tools narrows again: the host now sells research→finished-deliverable as a bundled subscription feature, with enterprise admin controls (Work Cloud/Work Local permissions, EKM) that speak directly to procurement.
2. The source layer hardens against unlicensed agents. Amazon's Sept 20 blockade of Meta's Muse — and its robots.txt blanket-refusal of 99 distinct named agents including GPTBot, ClaudeBot, PerplexityBot, and GoogleAgent-Mariner — pushes agent vendors toward the two compliant data paths: user-device-local operation, or licensed MCP backends (Agentic Browsing and the CFAA: The Ninth Circuit's Amazon v. Perplexity Ruling — and the Post-Legal Blockade of Agents, Parallel: Deep Research Sold as Per-Request Backend Infrastructure — with Audit-Grade Provenance and Shapley-Value Payouts to Publishers).
Net for positioning: the squeeze on standalone destination products is from both sides at once — hosts absorb research→deliverable workflows, and destinations gate raw web access2. The defensible layer remains what neither host nor destination wants to own: model-agnostic orchestration, evidence standards/audit, and MCP both-ways.3
-
An instance of Standalone deep research is dissolving into always-on agent runtimes. — The host folds research-to-deliverable into scheduled, trigger-driven agent runtime as a bundled subscription feature. ↩︎
-
An instance of Deep research bifurcates into destination products and per-request backends — survivors straddle both. — The bifurcation sharpens as hosts and gated sources close in on the destination middle from both directions. ↩︎
-
An instance of When models and retrieval commoditize, the orchestrator seat becomes the contested prize. — As hosts absorb research-to-deliverable and destinations gate raw access, the orchestration seat is the stratum left to contest. ↩︎