Cycle on Sep 21, 2026

I'll run this research cycle on the AI research-tool market. Starting by reviewing open threads and the current note set, since prior cycles already built a market map and pricing segmentation.

I have a well-developed note set — the market map, pricing segmentation, and per-player notes (Perplexity, Sacra, Elicit, Gemini, OpenAI, Scholar Labs) all exist. The HN map this cycle is mostly covered by prior notes (Panel, expertise-proxies). So this cycle's job: find what changed in the last ~24–48 hours across the market's players, and check whether any watches are close to firing (Perplexity's Sonar retirement is Sept 27). Running fresh searches now.

The searches surfaced two genuinely new market developments: (1) Perplexity's "Portable Computer" local-first agent now shipping on Windows via NVIDIA RTX + HP pre-loads, plus a Crusoe infrastructure deal — an extension of the local-first consumer-agent layer; and (2) Skillsync (YC W26), a Launch HN with real traction (65 pts, 58 comments) doing AI session portability across agents — a new interoperability layer adjacent to Hey, Lefty's workspace bet. Drilling into both.

The Perplexity blog is a substantive primary source — Portable Computer runs the "model, agent harness, orchestrator, and scheduler entirely on the Windows device" with local MCP servers and credit-free on-device work. That's Perplexity building the orchestrator layer at the desktop-OS level. The Skillsync story isn't in the cached scan, so I'll read its discussion directly, and also check the Panel thread for new comments since my Sept 16 note.

The Skillsync discussion is rich — founders articulating a portability thesis ("You should not get locked into one agent simply because you built up a conversation history + memory with that agent"), plus a substantive debate over whether the durable research artifact is a distilled note or the full transcript trajectory. The Panel thread also adds new detail (typed "Modules" for literature review, and a cluster of lookalike projects forming). Before writing, I'll ground the two remaining Perplexity claims — HP pre-load distribution and the Crusoe infrastructure deal — in their primary sources.

Both primary sources confirm the story: Perplexity now self-describes its flagship as "a model-agnostic orchestrator of AI agents across tools, files, chips, devices," ships it pre-loaded on HP workstations, and has locked in a full-lifecycle compute deal with Crusoe. I have everything needed. Writing three findings now: a revision of the Perplexity note, a new note on the session-portability layer (Skillsync), and a revision of the Panel/workspace-crowding note.

Perplexity note saved. The second call in that batch was sent empty — writing the Skillsync portability note and the Panel revision now, properly filled.

All three findings are persisted. Now registering the forward-looking tripwire the Skillsync thread surfaced: the portability layer dies if harness vendors lock down session-transcript access (there's already an opencode precedent in the thread).

The call failed because I omitted the required `title` parameter — retrying with it included.

Watch registered. Checking the thread list: all open threads are watches, and none fired this cycle — Perplexity's Sonar retirement (Sept 27) and OpenAI's post-mortem (Oct 1) are still ahead, and there was no Scholar Labs GA, Kimi internationalization, or Amazon v. Perplexity resolution in the past 24 hours. Nothing to resolve or record as an event. Ending the cycle.

Cycle complete. Here's what this research cycle found: ## What changed in the AI research-tool market (Sept 20–21) **1. Perplexity is now openly an orchestrator company — distributed through PC OEMs** *(revised `perplexity-deep-research-consumer-agent`)* - Its own new boilerplate: Perplexity Computer is "a model-agnostic orchestrator of AI agents across tools, files, chips, devices, and more," routing across "15+ models from OpenAI, Anthropic, and Google." - **Portable Computer for Windows** runs "the model, agent harness, orchestrator, and scheduler entirely on the Windows device," integrates via **local MCP servers**, and — the pricing signal — on-device work **consumes no credits**, decoupling long-running workflows from metered usage. Requires RTX GPUs with ≥24GB VRAM; Pro/Max tiers. - Distribution: **HP pre-loads** the app in the taskbar (ZBook Ultra G3a first), plus a Revit 2027 MCP integration; **Crusoe** deal covers training through managed inference. The compression threat to Hey, Lefty's layer now comes from a fourth direction (not just the host labs). **2. Session portability emerged as a new market layer** *(new note `agent-session-portability`)* - **Skillsync (YC W26)** launched with real traction (65 pts, 58 comments): a unified schema for agent sessions across Claude Code/Codex/Antigravity, so "the new agent gets the transcript: all the decisions in the trajectory, not a summary of it." Anti-lock-in pitch: "You should not get locked into one agent simply because you built up a conversation history + memory with that agent." - The thread's core debate — immutable transcript vs. distilled markdown note — is exactly the durable-artifact question for research tools; Hey, Lefty's notes+history architecture already bets on both layers. Registered a **tripwire watch**: harness vendors restricting transcript access (the opencode precedent) would kill this layer. **3. The agent-built-workspace mechanic is commoditizing** *(revised `agentic-research-workspaces-panel`)* - Panel added typed **"Modules"** (working literature-review module, early hypothesis generation) — but within days, four parallel open-source implementations surfaced (Nimbalyst, slices-ide, gptel). Pane-building can't be the moat; typed research workflows + evidence standards + source-layer MCP orchestration can. No open threads resolved; all existing watches remain pending (next dated catalyst: Perplexity Sonar endpoint retirement, Sept 27).