Agents

Sub-agents spawned for this project + when to use.

Used

• (none yet)

Candidates

• Explore — broad code search > 3 queries.
• general-purpose — open-ended multi-step research.
• frontend-design / impeccable:* — UI polish on _site/ output.

(log when invoked: date, agent, task, outcome)

• 2026-10-05: 4× general-purpose (parallel) — web discovery of new software/tools by slice (DCC/plugins, render/pipeline/VFX, engines/motion/CAD/sci-viz, AI). Each deduped vs a scratchpad catalog dump + verified live pages. Returned ~146 candidates; 151 added after merge. Gotcha: give agents the REAL vocab enums (I passed a guessed workflow list; 33 entries needed remap).
• 2026-10-05: Repo audit, 3 general-purpose agents in parallel (data quality, build/site code, CI/hygiene/docs). All read-only; scratch build in session scratchpad. Outcome: ranked findings delivered, nothing edited.
• 2026-10-05: Audit fix Phase 5, 7 general-purpose agents in parallel (rewrite descriptions A/B, backfill 151 new entries, §12 platform, §07 + §09 workflow, §12 misfile plan). Pattern that worked: agents write JSON patch files {key, set} to scratchpad only; orchestrator validates enums/style/length and applies via catalog.js (scratchpad p5-apply.js). Gotchas: one agent used nested tags:{} instead of dotted tags.platform keys (normalise before apply); apply moves LAST because moves change chunk keys; drop pricing_last_verified sourced from search snippets.
• 2026-10-31: /studios/ build. 5× research (vfx/VP/previs, animation, games/cinematics/outsourcing/mocap, archviz/product viz, XR/experiential/title) -> scratchpad research/out-*.json. 4× enrichment (SoM 500 -> fresh one-liners, disciplines, notable). 3× HQ verification. Shared BRIEF.md per job (rules + JSON shape + closed enums) kept prompts short. Gotchas: SoM ebook locations scrambled (Blur listed in Istanbul, Grind .it listed as Sofia) -> agents flagged ~35, auto site-text check flagged 185 for lookup; agents rename to site casing (BIGSTAR, Blur Studio); research facts sometimes from model memory, not fetched (games agent said so).
• 2026-10-05: 4× general-purpose (parallel) — mocap libraries, mocap datasets, Maya rigs (seeded from animationmethods.com/rigs.html), other-app rigs. JSON to scratchpad, orchestrator imported via scratchpad import.js (catalog.appendEntry). 166 candidates -> 160 new + 4 tools. Gotchas: §10 section slug is tools-pipeline-utilities; agents missed Mesh2Motion dupe (validate caught it); check-links.js --help is NOT a help flag, it starts a full scan.
• 2026-10-31: /studios/ v2. 8x general-purpose enrichment (95 rows each: founded, LinkedIn size band, 2D/3D re-tag, markets, with _src per fact) + 5x discovery (East Asia, S/SE Asia + Oceania, LatAm/MEA, Europe, NA + UK) + 2x India-sheet verification. Outcome: founded 169 -> 469 and size 84 -> 432 on existing rows; +478 studios. Gotcha: WebSearch has a 200-call cap per SESSION shared by every subagent (refills ~100/h). It ran out ~35 rows into the first agent; the rest fell back to WebFetch on guessed LinkedIn slugs (429s, wrong-company pages; agents rejected those on website/city mismatch) + Wikipedia + studio sites. Next time: fewer agents per wave, or tell agents to skip search and go straight to LinkedIn/Wikipedia fetches. Agents wrote progressive output files, which made partial dry-runs possible. Mainland China sites refuse non-CN fetches.
• 2026-11-03: 8x general-purpose in parallel. Catalog: web-tools 79, dcc-plugins 83, ai-cg 60, motion-vfx 74, assets 72 -> 367 added after cross-slice dedupe. Studios: asia 71, emea 67, americas+specialists 50 -> all 188 added. Shared BRIEF-catalog.md / BRIEF-studios.md in scratchpad, 18-WebSearch cap each: agents used 0-18; HN Algolia Show HN API, GitHub API/gh, itch.io and awesome-lists did most discovery. Reddit JSON blocked (403/HTML). LinkedIn curl fetches blocked by the permission system, so new studios mostly lack size. Overlaps between slices (MCP servers, Theatre.js, UVgami) caught by the importer.
• 2026-11-03: 10x general-purpose. Studios: archviz-europe 96, archviz-world 95, product-viz 68, specialist-cg 86, motion-niches 42 (no DNS in its sandbox: all from model memory, maintainer curl-probed, 31 kept). Games: na 289, east-asia 307, europe 903 (Wikidata via QLever, query.wikidata.org rate-limited), world 250. Plus 1 enrichment agent for 196 migrated rows (no network, memory only, 46 LOW). Gotchas: (1) some agent sandboxes had no DNS (router refuses bursts); main session network worked, so verify their output with scratchpad probe.js (curl status + final URL + title + spam/parked regex). (2) THREE agents put the user's email in a Wikidata User-Agent (Wikimedia UA policy asks for contact info). Briefs now carry a Privacy (HARD) rule; include it from the start next time. (3) Windows EPERM on rename when appendStudio rewrote the same chunk per row: batch writes per chunk + retry. (4) Probe regex false positives: "coming soon" on game sites (upcoming titles), "sedo" inside "mousedown".
• 2026-11-05: Studio round 4, 8x general-purpose in parallel (shared studios-r4/BRIEF.md, 15-WebSearch cap each, Privacy rule up front: no agent leaked contact info). prod-us 49, prod-europe 86, prod-asia 74, prod-specialist 50, prod-mea-latam 50, games-na 84, games-asia 59, games-emerging 62 -> 297 + 204 imported after host/name dedupe + main-session probe (Node fetch, curl recheck for TLS-chain fails). WebSearch use 0-15 per agent; QLever Wikidata + association member lists did most discovery. Gotchas: Wikipedia API 429s; agent DNS fails under burst (go low-concurrency); classifier denied a browser-UA curl retry and DoH bypass, so agents skip 403 sites; one agent's sandbox had no shell network (WebFetch only). Name-normalised dedupe has false positives (NExT Studios vs Next Games, NOMAD vs Nomad Games): override by host. Pixel×Pixel (pixelpixel.jp) done by hand in main session: festival sponsor/speaker lists are a good seed for regional motion studios.
• 2026-11-05: Studio gap fill, 8x general-purpose in parallel (shared research/BRIEF.md: rules, enums, style, 18-search cap, Privacy HARD rule). Studios: ai-interactive 60, capture-vp 53, medical-cgi 24, archviz 83, motion-ads 81; games: genre 72, ugc 50, regions 100. 500 passed import after dedupe + live probe. All used 13-18 searches; DNS flaky in several sandboxes (local filter 0.0.0.0), LinkedIn blocked, Wikipedia 429. Gotchas: probe spam regex parked and judi hit page code ("isParked", "Judicial"): check context before rejecting; Node fetch throws UNABLE_TO_VERIFY_LEAF_SIGNATURE on sites with missing intermediate certs that browsers accept; config "studios" agents return SaaS platforms (drop). No email leaks this round.
• 2026-11-06: Country fill, 8x general-purpose by region (both directories each). Session restarted ~10 min in: only 23 rows saved; all 8 resumed via SendMessage from transcripts and finished (93/72/99/68/43/39/52/67). 524 kept. Gotchas: resume works, partial files were kept and appended; small-country sources are JS-only association lists (GDAP, AINAKI, TACGA, BGDA API denied by permission system); DeVuego useful for LatAm; agents flag _new_country reliably, so check unknown countries across all out files before import.
• 2026-11-06: Round 5, 9x general-purpose by domain, event-sourced (shared research/BRIEF.md, 15-search cap, Privacy rule). Raw out: vfx 70, animation 65, interactive-xr 65, motion-ads 44, niche-craft 19, viz 47, games-west 66, games-world 56, game-services 52 -> 471 imported. WebSearch use 1-14 per agent: event sites + Wikipedia ceremony pages did the work. Best sources: public JSON behind event sites (FMX companies, ADFEST archive, BIG Festival games API, Berlin MVA WP API), Cartoon Forum/Movie producer lists, Venice Immersive work pages (producer sites), CGarchitect gallery crawl. Gotchas: many award sites JS-only (CICLOPE, AICP, Clio, Promax, Thea, Telly); Wikipedia VES/HPA pages credit people not vendors; classifier blocked repeat API paging (BIG) and cert-check bypass; niche-craft agent wasted calls verifying rows already listed (dedupe before verifying). No email leaks.
• 2026-11-08/09: News discovery, 8x general-purpose by slice (shared scratchpad news/BRIEF.md + SUBSECTIONS.md + catalog.tsv dump, 18-search cap, Privacy rule). Raw: ai-gen 77, art-design 78, assets-learning 67, dcc-render 70, gamedev 68, hardware-xr 56, research-scan 81, vfx-motion 63 + 34 update proposals -> 548 imported, 26 updates applied. WebSearch use 3-13 per agent (~68 total); best sources: CG Channel weekly roundups, gamefromscratch.com, radiancefields.com sitemap, ComfyUI blog RSS, HN Algolia, Digital Production archive, CGPress feed. Gotchas: unauthenticated GitHub API rate limit shared across agents (fall back to HTML/raw LICENSE); browser-UA curl blocked by classifier (agents skipped bot-walled sites); news thin for art/assets slices, so agents filled older gaps; one agent ran rm -f b*.json in the shared folder (tell agents to use a private subfolder); several update proposals overlapped across slices (Megascans x2, Weavy x2, Cavalry x2).
• 2026-11-09: Studio round 6, Europe by country. 14x general-purpose (UK, DE, FR, ES+PT, IT+GR+MT, Benelux, AT+CH+IE, SE+NO+IS, DK+FI, PL, CZ+SK+HU, Baltics+UA+BY+MD, Balkans+RO+BG, TR+CY+Caucasus), shared research/BRIEF.md, 12-search cap each (agents used 3-12, ~100 total, no cap hit), private work-<slice>/ folders. Raw 45-100 per agent, 935 imported. Best sources: association member JSON APIs + Steam appdetails for release checks; CEE Animation REST API. Gotchas: one agent saved downloaded pages into the repo root (moved out; tell agents to fetch into work-<slice>/ only); agents flagged existing-row errors (Pikkukala, Nebula, Hangar 13): useful, check before acting (2 of 3 were already fine); spam regex again hit 'Judith'/'prejudice' (judi), game titles and ad clients; QLever moved to qlever.dev and 429s under load. No email leaks. Runs took 27-64 min each.
• 2026-11-09: Studio round 7, Asia-Pacific by country. 13x general-purpose (Japan prod, Japan games, India prod, India games, Australia, NZ+Pacific, Korea, China, TW+HK+Macau, SG+MY+Brunei, ID+PH, VN+TH+mainland SEA, South+Central Asia), shared research/BRIEF.md, 12-search cap, private work-<slice>/ folders. Raw 20-110 per agent, 712 imported. Runs 26-90 min. Best sources: Wikidata/QLever, CGWORLD studio DB, association WP/JSON APIs (BitSummit, AnimationXpress, BASIS, NZGDA, HKVFX, MDEC, SGGA), Steam curator lists, Google Play dev addresses. Gotchas: (1) one agent used the user's connected Tracxn MCP for leads (brief had no rule; now forbidden: public web only). (2) one agent tried to extract an API key from a JS bundle (permission system denied; now forbidden in brief). (3) TW agent found a film-commission API exposing contact passwords, deleted them; main session confirmed none left. (4) leads handed between split agents fell through (4 India game devs). (5) China agent used Steam search URLs as studio url (dropped; /developer/ pages kept). No email leaks.
• 2026-11-09: Studio round 8, South America + Africa by country. 12x general-purpose (Brazil prod, Brazil games, AR+UY+PY, CL+PE+BO, CO+EC+VE+Guianas, South Africa, southern Africa + islands, Nigeria + Anglophone West, Francophone West/Central, East Africa, Egypt + Sudan, Maghreb), shared research/BRIEF.md, 12-search cap, private work-<slice>/ folders. Raw 4-80 per agent (328 total), 319 imported. Runs 29-69 min. Africa slices hit half or less of target (studios with Facebook-only or no site). Best sources: gamesindustryafrica.com WP API, Africa Games Week catalogue, ABCA PDF, Jogos Brasileiros Steam curator, BIG Festival API, ADVA/Animachi/Morocco Gaming Expo WP APIs. Gotchas: QLever 429 for most agents; one agent tried a JS-bundle API again (denied); 3 helper scripts left in research root; one agent ran 13 WebSearches (cap 12); several agents deleted PDFs/Play pages holding personal contact data (correct). No email leaks.