# Trending Scout

> Use when: vendor/agent landscape scan, 'what's new', trending repos, weekly update check. Filtered vs known memos — NOT deep topic research (/research).

- Skill: `markusstrasser/trending-scout` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add markusstrasser/trending-scout`
- Raw SKILL.md: https://api.skillmd.com/api/skills/markusstrasser/trending-scout/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: markusstrasser (https://skillmd.com/u/markusstrasser)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/markusstrasser/trending-scout

---


# Trending Scout

Scan for genuinely new agent/AI developments, filtered against the known landscape. The goal: surface what's changed since last check, not rehash what's already in research memos.

## Core Principle

**Knowledge-diff, not news aggregation.** Every finding must answer: "What does this change about what we know or how we build?" Repos with 50K stars that don't affect our architecture are less interesting than a 200-star tool that solves a problem we have.

**Dev sources only.** Prioritize: API docs, dev docs, changelogs, release notes, trending repos (code you can read), "awesome-X" / "best of" lists. Deprioritize: tech journalism, announcement blog posts, hype pieces, product launches without code. The signal is in changelogs and READMEs, not TechCrunch.

## Input Handling

Accept optional focus:
- **No argument / "all"** — full scan across all categories
- **Vendor name** ("anthropic", "openai", "google") — vendor-focused scan
- **Topic** ("mcp", "frameworks", "benchmarks") — topic-focused scan
- **"weekly"** — full scan formatted for orchestrator pipeline output

## Phase 0: Establish Baseline

Before searching, load the known landscape so you can filter against it.

1. **Read the research index** — `.claude/rules/research-index.md` in meta (always auto-loaded if in meta). This tells you what topics have existing memos.

2. **Check the most recent trending-scout memo** — `research/trending-scout-*.md` in meta. This is the "last known state." Note dates and what was already evaluated.

3. **Check vendor-versions baseline** — run `uv run python3 ${CLAUDE_SKILL_DIR}/scripts/vendor-versions.py` for current SDK/CLI versions. This catches version bumps that searches might miss.

4. **Read tracked agent entities** — glob `analysis/agent-entities/*.md` in meta. Each file has `last_refreshed` frontmatter and a structured Current State section (version, pricing, context window, transport). Treat these as the canonical per-entity baseline. **Scout runs are the sole refresher** — the `agent-entity-refresh` pipeline died with the orchestrator (eradicated 2026-06-07; files sat as unpopulated seeds until the 2026-06-11 scout populated them). Update entity files per Phase 3 whenever the scan covers their vendor; note any entity the scan did NOT cover in the search log. See `decisions/2026-04-10-agent-entity-surveillance.md` for the original architecture rationale.

5. **Note the time window** — default is "since last scout memo date" or "past 7 days" if no prior memo exists.

## Phase 1: Multi-Source Parallel Scan

Dispatch parallel searches across sources and categories. Use subagents for parallelism when available (4+ independent search axes = always delegate).

### Search Matrix

| Category | Exa Query | Brave Query | Extras |
|----------|-----------|-------------|--------|
| **Anthropic** | "Anthropic Claude changelog" + date filter | "Claude Code release notes OR changelog" | WebFetch changelog + SDK releases (see Anthropic deep-check below) |
| **OpenAI** | "OpenAI Codex CLI changelog" + date filter | "OpenAI API changelog OR SDK release" | Check github.com/openai releases |
| **Google** | "Gemini API OR google-genai SDK release" + date filter | "Gemini API update OR Antigravity agy CLI" | Check github.com/google-gemini releases (free Gemini CLI retired 2026-05-31 → Antigravity `agy`) |
| **Frameworks** | "AI agent framework" + date filter | "agent framework release 2026" | Check github trending |
| **MCP** | "MCP server model context protocol" + date filter | "new MCP server" | Check github.com/modelcontextprotocol |
| **Repos/Lists** | "awesome AI agents" + date filter | "trending AI repos github" | GitHub trending via Exa site filter |
| **Tools** | "AI coding tool SDK" + date filter | "cursor OR windsurf OR coding agent SDK changelog" | — |
| **Research** | arxiv "LLM agent" + date filter | — | alphaXiv trending (`WebFetch alphaxiv.org/explore`), S2 search |

### Anthropic Deep-Check (when focus is "anthropic" or "all")

This is the highest-signal category for our infrastructure. Don't just keyword-search — hit these sources directly:

1. **Claude Code changelog** — `WebFetch` the official changelog at `code.claude.com/docs/en/changelog`. Compare version numbers against `vendor-versions.py` output. Focus on: new hook events, new frontmatter fields, new CLI flags, SDK changes.

2. **Agent SDK releases** — Check `github.com/anthropics/claude-agent-sdk-typescript/releases` and `github.com/anthropics/claude-agent-sdk-python/releases` via Exa site-filtered search. Note new `query()` options, hook input fields, tool changes.

3. **Cookbook new patterns** — `web_search_advanced_exa` with `includeDomains: ["platform.claude.com"]` and date filter. Look for new notebook patterns (agent architectures, MCP patterns, tool use patterns).

4. **GitHub issues for features we track** — Search `github.com/anthropics/claude-code/issues` for issues we're watching (tool output compression #32105, Agent Teams stabilization). Check if closed/merged.

5. **Compare against deferred items** — Read `research/claude-code-native-features-deferred.md` in meta. Check if any "trigger to revisit" conditions are now met.

**Output for Anthropic category:** Version delta (old → new), new features categorized as (adopt now / evaluate / defer), and update to deferred-items memo if triggers are met.

### Source-Specific Techniques

**Exa** (primary — semantic search, good for discovering things you wouldn't keyword-match):
```
web_search_advanced_exa:
  query: "<category query>"
  numResults: 10
  startPublishedDate: "<window_start in YYYY-MM-DD>"
  type: "auto"
```

**Brave** (secondary — independent index, good for triangulation):
```
brave_web_search:
  query: "<category query>"
  count: 10
  freshness: "pw"  # past week
```

**arxiv** (for research papers):
```
search_arxiv:
  query: "AI agent <topic>"
  max_results: 5
```

**alphaXiv** (trending research — community-curated signal, higher SNR than raw arxiv):
```
WebFetch:
  url: "https://www.alphaxiv.org/explore"
  prompt: "Extract all trending papers: title, arxiv ID, view count, summary"
```
Weekly digest also arrives via email — check for alphaXiv newsletter content if user provides it. View counts indicate community attention; papers with 1K+ views in agent/LLM categories are high-signal.

**GitHub trending** (for repos — use Exa with site filter):
```
web_search_advanced_exa:
  query: "AI agent tool"
  numResults: 10
  includeDomains: ["github.com"]
  startPublishedDate: "<window_start>"
```

### Parallel Dispatch Pattern

When subagents are available, dispatch one per vendor category:
- Agent 1: Anthropic (Exa + Brave + anthropic.com)
- Agent 2: OpenAI (Exa + Brave)
- Agent 3: Google (Exa + Brave)
- Agent 4: Ecosystem (frameworks + MCP + tools via Exa + Brave + GitHub)
- Agent 5: Research papers (alphaXiv trending + arxiv + S2)

Each agent returns: `[{title, url, date, one_line_summary, why_relevant}]`

When running without subagents, execute searches sequentially but batch by source (all Exa queries together, then all Brave queries).

## Phase 2: Deduplicate and Filter

### Against Known Landscape

For each finding, check:
1. **Already in a research memo?** — grep research/ for the repo name, tool name, or key concept
2. **Already in CLAUDE.md or rules?** — might be documented as infrastructure we use
3. **Version bump of known tool?** — compare against vendor-versions output. Note the delta but don't treat as "new"

### Quality Filters

Drop findings that are:
- **Tech news / hype** — journalist write-ups, "revolutionary AI agent" blog posts, product launch PR. If there's no code, changelog, or API doc behind it, skip it.
- **Pre-alpha / concept only** — README-only repos, no releases
- **Irrelevant domain** — AI developments that don't touch agent infrastructure (e.g., image generation models, unless they have agent implications)
- **Duplicates** — same finding from multiple sources (keep the most informative source)
- **No code to read** — if you can't look at a repo, SDK, or API surface, it's not actionable

### Relevance Scoring

Rate each surviving finding on two axes:
- **Value**: How much does this change what we know or how we build? (1-5)
- **Maintenance**: What would it cost to adopt/integrate? (1-5, lower is better)

Keep findings where Value > Maintenance, or where Value = 5 regardless of maintenance.

## Phase 3: Output

### Memo Format

Write to `research/trending-scout-YYYY-MM-DD.md` in meta:

```markdown
# Trending Scout — YYYY-MM-DD

**Date:** YYYY-MM-DD
**Window:** [start] to [end]
**Sources:** Exa, Brave, arxiv, GitHub
**Findings:** N new, M version bumps, K already known (filtered)

---

## New Findings (ranked by value - maintenance)

### 1. [Name]

| Field | Content |
|-------|---------|
| Source | [URL] (stars if GitHub, citations if paper) |
| What it does | One paragraph |
| Why relevant | How this relates to our infrastructure |
| Integration path | What we'd do with this (adopt / extract pattern / watch / ignore) |
| Current overlap | What we already have that's similar |
| Maintenance cost | Ongoing drag if adopted |
| Verdict | **Adopt** / **Extract pattern** / **Watch** / **Ignore** |

### 2. ...

## Version Bumps

| Tool | Previous | Current | Notable Changes |
|------|----------|---------|-----------------|
| ... | ... | ... | ... |

## Already Known (filtered out)

Brief list of things found in search that we already track — confirms coverage, no action needed.

## Search Log

What was searched, what returned useful results, what didn't. Helps calibrate future scans.
```

### Entity File Updates

When a finding reveals state that is fresher than what's in `analysis/agent-entities/<entity>.md` — e.g., the scan catches a version bump or pricing change before the next refresh-pipeline run — edit the entity file in place:

- update Current State fields
- prepend a dated entry to Recent Changes with `[trending-scout]` tag and a source URL
- bump `last_refreshed` in the frontmatter to today's date
- do NOT edit Monitoring Triggers or Sources sections unless a trigger fired (in which case add `**TRIGGER FIRED**` at the top of Recent Changes)

Don't create new entity files during a scout run. If a scan surfaces a new tool worth tracking, note it in the memo's "New Findings" section with a suggested entity file to create on the next refresh cycle — entity file creation is a deliberate act, not a scan side-effect.

### Pipeline Output

When invoked as `--weekly` (or dispatched from `/observe maintain`'s weekly frontier rotation), also:
1. Update the research index in `.claude/rules/research-index.md` if new memo warrants a permanent entry
2. Commit the memo and any entity file updates: `[research] Trending scout — N new findings, window YYYY-MM-DD to YYYY-MM-DD`
3. **Route findings into the consumption loop — NOW, at source (immediacy).** A memo is
   generation; the loop drains `improvement-log.md`, NOT `research/`. Routing here, in the
   same session that found them, is strictly better than waiting for the weekly harvest to
   re-derive them — you have full context. **Follow the one canonical finding-routing protocol
   defined in `/observe` harvest Phase 2f** (enforced by `just orphan-findings`); do NOT restate
   its rules here — single source, so the two never drift (constitution principle 9). In short:
   Adopt/Evaluate → `[ ]` with the finding title verbatim; no-live-items memo → one
   `RECONCILIATION:` entry; Watch/Ignore stay memo-only. Harvest 2f is the weekly backstop for
   anything missed here; `just orphan-findings` measures the gap.

## Loop Integration

This is a **worker** skill, not a loop conductor. It runs on a weekly cadence as one of the
frontier-scan rotation rows in `/observe maintain` (alternating with `/observe architecture`),
or by hand as `/trending-scout weekly`. (The standalone `orchestrator` pipeline that used to
schedule it was retired 2026-06-12 — `/observe maintain` is the single conductor now.)

## Edge Cases

- **No new findings**: Write the memo anyway — "no new findings" is a valid data point that confirms the landscape is stable. List what was searched.
- **Overwhelming volume**: Cap at 10 findings per scan. Rank aggressively. Link to sources for the rest.
- **Stale baseline**: If no prior trending-scout memo exists, the first run is a broader landscape scan. Subsequent runs are diffs.
- **Rate limits**: If Exa or Brave rate-limits mid-scan, note which categories were incomplete in the search log.