Cycle on Oct 8, 2026

I'll start by reviewing the open threads and existing notes for this topic, then drill into the HN stories that touch the AI research tools market — most notably Panel (a research workspace Show HN) and the "AI is breaking our proxies for expertise" piece.

The topic has a rich existing note set. Two HN stories in the current map directly touch this market — Panel (already tracked under `agentic-research-workspaces-panel`, last revised Oct 2) and the "AI is breaking our proxies for expertise" essay (tracked under `ai-expertise-proxies-goodhart`, last revised Oct 4). I'll check both for new material since those revisions, and pre-check the Gemini free-tier cut that fires tomorrow (Oct 9) per an existing watch.

Both tracked HN stories are static since the last revisions (Panel: same 4 comments as the Oct 2 note; Goedecke essay: same 55 pts / 24 comments as the Oct 4 note). The Gemini Oct 9 cut is confirmed on schedule by multiple outlets — that watch fires tomorrow, not today. Let me sweep for anything genuinely new this week in the research-tool market, and trace the "AI writing smell" comments from the HN map (a distinct expertise-proxy signal) back to their parent story.

Both tracked HN threads are static, so the live signal is in the wider web sweep: Perplexity's October wave (Decider API + open-weights model, Amex "Skills" distribution, Rings connector, Vera-silicon vertical integration), Microsoft's Oct 7 move giving Copilot local file context and OS-wide actions (directly relevant to the local-first layer), and a new AI-text "watermark stripper" (TextGrain) that ties into the expertise-proxies thread. Let me verify each at primary or solid secondary sources.

Right — I have verified evidence for three distinct findings; now I need to write them up and close the cycle properly. Two quick verification fetches first (the Hugging Face page for Perplexity's open-weights decider model, and the Rings AI partnership release), then I'll write the notes.

All evidence is verified. Writing the three findings now — one new subject (Microsoft's local agent stack) and two upserts of existing slugs (Perplexity's October wave; the expertise-proxies laundering layer).

All three findings are persisted. Now registering the forward-looking items this cycle surfaced: two watches (Microsoft's local-agent rollout as a tripwire on the local-first window; the Surface Laptop Ultra ship date) and one open question (cheap decision models as the grader substrate for research-agent evals).

Cycle complete. Here's what this research cycle found and persisted: ## Findings written (3) **1. Microsoft's local agent stack — new note** (`microsoft-copilot-local-agent-stack`) Microsoft's Oct 7 "hybrid intelligence" announcement gives Copilot on Copilot+ PCs local file context, OS-wide actions, and on-device models (MAI Code 1.1 Flash at 3-bit/256K context; DeepSeek V4 Flash 284B on RTX Spark), and **MXC — its agent-containment layer — went GA the same day**. OpenClaw and Codex already support it; Perplexity and Meta's Muse are queued. This compresses Hey, Lefty's local-first differentiation window from the *OS side*: "runs locally" stops being a differentiator when the default OS ships it natively, and MXC becomes a new compliance gate for any agent acting on Windows. **2. Perplexity's October wave — updated** (`perplexity-deep-research-consumer-agent`) - **Decisions API**: a typed "judgment primitive" (`noul`/`choice`/`score` probabilistic answers) at **$0.02/M input tokens, free output**, with open weights (`pplx-decider-v1-27b`, Qwen3.8-27B fine-tune) on Hugging Face — self-graded benchmarks included. - **Amex deal**: 10 pre-built "Skills" for Perplexity Computer distributed through a card-issuer channel ($125 credit + 1,000 bonus Computer Credits), Enterprise-only. - **Rings connector**: relationship-graph context, with the PR describing Perplexity as "a model-agnostic orchestrator" — the same positioning slot the market map assigns to Hey, Lefty. **3. The expertise-proxy laundering layer — updated** (`ai-expertise-proxies-goodhart`) TextGrain Remover (Oct 7) offers free stripping of AI watermarks and statistical token signatures while preserving facts/citations — proxy erosion now runs both directions: humans falsely dismissed for "AI smell," AI work laundered past detection. Provenance, not detection, is the durable trust axis. ## Also checked - **Panel** and the **Goedecke essay** HN threads: both static since the last note revisions — no updates needed. - **Gemini free-tier cut (Oct 9)**: confirmed on schedule by The Verge/Engadget/9to5Google — the existing watch fires *tomorrow*; next cycle should verify the actual effect. ## Registered for future cycles - Watch: Microsoft hybrid Copilot EEA availability / GA beyond preview - Watch: Surface Laptop Ultra ships Oct 16 (120B+ param local models) - Open question: will cheap decision models become the standard grader substrate for research-agent self-evals?