1. Round 1discover14 of 14 backprofile-pilot5 of 5 back

   Agent #2 planned it. Started 1 h 8 min after the job was posted. 2 of its items were sent back for a funding_note over 1,000 characters.

   what discover was asked

   Find every company in this segment that builds humanoid robots: {{segment}}. Starting points: {{hints}}. Then search the open web (company sites, trade press such as The Robot Report and IEEE Spectrum, funding news, exhibitor lists from CES, ICRA, the World Robot Conference in Beijing, Humanoids Summit, investor lists such as the Morgan Stanley Humanoid 100 and The Humanoid Hub) until new searches stop turning up new names. A company counts when it designs and builds its own general-purpose humanoid robot: a torso with two arms and a head on legs or a wheeled base, meant for sale, deployment or research use, as a product or a publicly shown prototype. Automaker and big-tech divisions, listed companies and startups all count. Do not count pure component makers (actuators, hands, sensors), software-only or foundation-model companies without their own robot, universities and research institutes, hobby projects, single arms, exoskeletons, or small toy or education humanoids under 60 cm unless the company also builds a larger one; list such doubtful cases with include set to "borderline" and a one-line reason instead of dropping them. Use the company's official English name without legal suffix (Inc, Ltd, GmbH, Co) as its canonical name; for Chinese companies use the English name the company itself uses and put the Chinese name in note. Open every source_url you cite and confirm it shows the company building a humanoid; a company you cannot source is left out. Result: one JSON object and nothing else: {"segment": string, "companies": [ {"company": string, "hq_country": string (English short country name), "website": string, "humanoid_product": string (robot name and one phrase on what it is), "founded_year": integer or null, "source_url": string, "include": "yes" or "borderline", "note": string} ], "sources_checked": [ strings, the URLs you used ]}. Up to 120 companies per result.

   what profile-pilot was asked

   Profile one humanoid-robot company: {{company}} ({{hint}}). Use only public web sources and give a URL for every figure: the company's own site, filings (SEC, HKEX, SSE), press releases, then reputable press (Reuters, Bloomberg, FT, TechCrunch, The Robot Report, IEEE Spectrum, 36Kr for China). Fill exactly one row of the table you are given, keyed by company = "{{company}}". what_it_builds: one or two sentences on its humanoid line and intended use. hq_country: English short country name of the headquarters. founded_year: year of incorporation. funding_usd: total disclosed equity capital raised to date in US dollars, the sum of the rounds you list in funding_note, each with its date, amount and URL; convert other currencies at the rate of the announcement date and state the rate; exclude debt, grants and valuation. funding_basis: one of "disclosed rounds", "public company", "parent funded", "undisclosed". If the company is listed or a division and discloses no humanoid-specific raise, leave funding_usd empty, set funding_basis accordingly and give IPO proceeds or parent in funding_note. latest_funding_date: date of the newest round. valuation_usd: the latest reported post-money valuation, else empty. latest_model: the newest humanoid model publicly announced, with latest_model_year and latest_model_source. units_shipped: only a number the company, a filing or a named reputable report has published (delivered, sold or deployed units; say which, with the date, in units_shipped_note); never an estimate; empty when no public figure exists. status: "active", "acquired" or "defunct". as_of: today's date. Every *_source column is one URL that actually shows the figure in the same row; open each URL to confirm; a figure without a source is left empty rather than guessed. Put the row and nothing else in your JSON result.

   the whole memo

   EPISODE 1 (2026-11-03). Fresh goal: 1000 items, 100 workers, 15 episodes, nothing run yet. Plan in four phases: (1) DISCOVER the universe of humanoid-robot companies, (2) PROFILE each company into one table row (the goal's table), (3) VERIFY the figures the ranking stands on, (4) FINISH and write the table plus ranking. DEFINITION used in every brief (keep it stable): a company counts when it designs and builds its own general-purpose humanoid robot (torso, two arms, head, on legs or a wheeled base) meant for sale, deployment or research use, as a product or a publicly shown prototype; automaker/big-tech divisions, listed companies and startups all count. Excluded: component makers (actuators, hands, sensors), software-only/foundation-model firms with no robot of their own, universities/institutes, hobby projects, single arms, exoskeletons, small toy/education humanoids under 60 cm. Doubtful cases are listed with include=borderline and a reason, so the orchestrator decides, not the worker. Canonical name = official English name without legal suffix; Chinese companies by the English name they use themselves. SPAWNED THIS EPISODE (19 items): batch `discover`, 14 research items, json output, merge groupSize 7 (so merges merge-01 and merge-02). One item per region or source type: usa, canada-latam, china-startups, china-corporates, japan, korea, europe-west, europe-other, india-mideast-africa, asia-other-oceania, trackers-lists, automakers-bigtech, funding-news-2025-2026, open-hardware-spinouts. Overlap between items is intended; the merge dedups. Batch `profile-pilot`: 5 table items (figure-ai, agility-robotics, unitree, 1x-technologies, ubtech-robotics) that test the profile brief and the table shape on a US startup, two US/Norway startups, a Chinese private firm and a HK-listed company (public-company funding handling). Table columns: company(key), hq_country, website, status, what_it_builds, what_it_builds_source, founded_year, founded_source, funding_usd (USD, totalled, optional), funding_basis, funding_source, funding_note, latest_funding_date, valuation_usd, latest_model, latest_model_year, latest_model_source, units_shipped (optional), units_shipped_source, units_shipped_note, as_of. Reads requested: every discover item, both merges, all 5 pilot rows. Wake when both batches settle or after 180 minutes. EPISODE 2 SHOULD: (a) Check goal.json for refusals. If `profile-pilot` was refused because the plane does not keep datasets (table unsupported), re-spawn the profile batch as plain json output with the same fields and merge groupSize 20 into a table. If `discover` paused in canary, read its failures and mend the brief (likeliest problems: JSON too big -> raise maxBytes or cut per-item cap; dead source URLs -> demand the worker open each URL). (b) Read discover/merge-01 and merge-02 (and the items if merges are thin), build the canonical company list: include=yes rows plus borderline rows that fit the definition; dedup by name/website. Expect roughly 120-250 companies (China alone has 60+; the Morgan Stanley Humanoid 100 and The Humanoid Hub lists are good cross-checks). (c) Read the 5 pilot rows: check sources are real URLs, funding sums match the listed rounds, units_shipped is only ever a published figure. Mend the brief where the rows show a weakness, then spawn `profiles-a` (new name) with the full company list, one item per company, the same table shape (column names must match exactly so the write-up can union datasets), groupSize ~20, maxRows 1. Keep ~400 items of budget for later verification and gap-filling. (d) If the discovery list is clearly incomplete in one region, spawn a second targeted discover batch (new name, e.g. `discover-2`) alongside. EPISODE 3-5 SHOULD: read profiles failures and disputed cells; re-run failed items with a mended brief via itemsFrom; spawn `verify-top` with replicas {size:3} for the ~30 best-funded companies (funding_usd, valuation_usd, units_shipped, latest_model) since the ranking rests on those; spawn a `signals` batch for the top ~25 collecting ranking evidence not in the table (paid deployments and named customers, production capacity/targets, headcount, notable demos, autonomy claims vs teleoperation) with sources, output markdown, so the ranking has reasons. Then finish (target episode 6-8) and write: one table keyed by company with total funding from the dataset totals, ranking with criteria stated (funding, units shipped/deployed, latest model maturity, customers, manufacturing scale), and a list of what is missing (companies with no disclosed funding, disputed cells). BUDGET NOTE: 19 of 1000 items used after this episode. Do not spend more than ~350 on profiles, ~120 on replicas/verification, ~60 on signals; keep a reserve for re-runs.

   3 steps

   1. Orchestrate 001Agent #2made 1 file

      All checks passed. The git status call failed only because the sandbox blocks reading the global git config, which does not affect the deliverable. The file is written and left untracked for the daemon to upload.

      Outcome: episode 1 decisions are written to artifacts/goal/episodes/001/decisions.json as one JSON object in the reference shape.

      What it decides

      • Batch discover (14 research items, json, merged in groups of 7). One item per region or source type: USA, Canada and Latin America, China startups, China listed and corporate, Japan, Korea, Western Europe, other Europe, India/Middle East/Africa, other Asia and Oceania, public trackers and indexes, automaker and big-tech programmes, funding news since January 2025, and open-hardware and spin-outs. Each item returns a JSON list of companies with country, website, product, founding year, a verified source URL and an include flag. The brief fixes one definition of "builds humanoid robots" and routes doubtful cases to a borderline flag rather than letting workers drop them.
      • Batch profile-pilot (5 table items): Figure AI, Agility Robotics, Unitree, 1X Technologies and UBTech Robotics. This tests the per-company table brief on a mix of US, Norwegian, Chinese private and Hong Kong listed companies before it is spent on hundreds. The table has 21 columns covering the goal's fields plus a source URL for each figure, funding basis and notes, with funding and units shipped as totals.
      • Reads for the next episode: all 14 discover items, both merges, and the 5 pilot rows. Wake when both batches settle or after 180 minutes. Not finished.

      Checks run locally

      Check	Result
      Notes length	4,780 of 8,000 chars
      Longest filled brief	2,385 of 3,000 chars
      Items spent	19 of 1,000
      Reads requested	21 of 30
      Column names, types, key, totals	valid

      Next episode's plan is in the notes: build the canonical company list from the merges, mend the profile brief from the pilot rows, spawn the full profile batch under a new name with the identical column set, then verify the top companies with replicas and gather ranking evidence before finishing around episode 6 to 8. The notes also say what to do if the plane refuses the table batch.

      ran onclaude · claude-fable-5-1 · 6 turns · 3m 42s · 98 in · 16.3K out · 172.7K cached

      submission3bd6a3cff21f0f707486a70fd5338204dd413670cbaf30e6970f48d88b46a92f

      device468e82a89b9bfe188ef46b055e2ea554bde626cc162630c20effb0fcc7e13fed

      started from0243d7da4a4337ae8b16bcdf15bb4ead736fd68f

      bundlenone

      verifiedrebuilt and matched · verifier 0.1.0 ·

      made · 1 file

      artifacts/goal/episodes/001/decisions.json · 19 KB

      may write

      artifacts/goal/episodes/001/**
   2. Merge b01 01Agent #1576made 1 file

      {"output":"artifacts/goal/batches/discover/merges/01/result.md","companies":296,"bytes":63997,"validation":"JSON parsing and merge checks passed"}

      ran oncodex · 4 turns · 6m 20s · 78.6K in · 13K out · 990.1K cached

      submission022d96b5a0aead1287546668adc866310710ee0a0b3371ba72cad788d6f75281

      device6412d5c37aa6592f06eb01311ad33f383bc15f6fc3417f0c7fcb5d0a95fb2903

      started from0243d7da4a4337ae8b16bcdf15bb4ead736fd68f

      bundlenone

      verifiedrebuilt and matched · verifier 0.1.0 ·

      made · 1 file

      artifacts/goal/batches/discover/merges/01/result.md · 64 KB

      may write

      artifacts/goal/batches/discover/merges/01/**
   3. Merge b01 02Agent #858made 1 file

      Created and validated result.md.

      • 229 deduplicated companies
      • 7 source items merged
      • Sorted and unique
      • Valid JSON
      • 62,226 bytes (under 64,000)

      ran oncodex · 4 turns · 1m 39s · 73.9K in · 12.4K out · 797.1K cached

      submissiondf42a856f6153c6df9f4209dd85b7aa726883b3195490c1babb61f40b6667a43

      device68fd8dd6450ac70c17815a4b8406630872b015196d49e05b17fe03c8fb319fe6

      started from0243d7da4a4337ae8b16bcdf15bb4ead736fd68f

      bundlenone

      verifiedrebuilt and matched · verifier 0.1.0 ·

      made · 1 file

      artifacts/goal/batches/discover/merges/02/result.md · 62 KB

      may write

      artifacts/goal/batches/discover/merges/02/**

   discover, item by itemprofile-pilot, item by item

2. Round 2profiles-a259 of 275 back, 16 cancelleddiscover-25 of 5 back

   Agent #785 planned it, 2 attempts. 136 of its items were sent back for a funding_note over 1,000 characters.

   what profiles-a was asked

   Profile one humanoid-robot company: {{company}} ({{hint}}). Fill exactly one row of the table you are given, with company = "{{company}}" written exactly so. Open every URL you cite and cite only pages you opened. Prefer, in order: the company's own site and press releases; regulator filings (SEC, HKEX, SSE, SZSE, Companies House); reputable press (Reuters, Bloomberg, FT, Nikkei, TechCrunch, The Robot Report, IEEE Spectrum, Yicai, Caixin, 36Kr). Never cite Wikipedia, Crunchbase, Tracxn, PitchBook or CB Insights pages. If a page is blocked, find another source for the same fact rather than dropping the fact. hq_country: English short name of the headquarters country (United States, China, South Korea, Hong Kong, Taiwan...). website: home page URL. status: one of active, acquired, defunct, division (a unit of a larger group: name the parent in what_it_builds). what_it_builds: 1-2 sentences on its humanoid line and intended use, with what_it_builds_source. founded_year: year the company was incorporated, with founded_source (company site, filing or press); empty only if no source states it, and say so in funding_note. funding_usd: whole US dollars, the sum of every disclosed equity raise you can source. Private company: every round with a stated amount; a round reported as 'over X' counts as X. Listed company: pre-IPO rounds plus IPO gross proceeds plus later share placements. Division of a group: leave empty and set funding_basis = corporate division. Rounds reported by reputable press count even without company confirmation; mark them 'press-reported' in funding_note. Convert other currencies at the rate the source states, else at the announcement-date rate, and say which. funding_basis: disclosed rounds | public company | corporate division | none disclosed. funding_note: every round you summed as date, amount, round name and URL; every reported round you excluded and why; the valuation's date. funding_source: the URL of the largest round summed. latest_funding_date: date of the newest round summed. valuation_usd: latest reported post-money valuation or market capitalisation, empty if none. latest_model: the newest humanoid model publicly shown, latest_model_year: the year it was unveiled, latest_model_source. units_shipped: a published count of humanoid units delivered, sold or shipped, cumulative if available, else the latest full-year or half-year figure; never orders, pre-orders, capacity, targets or analyst estimates. units_shipped_note says what period and which products the count covers; units_shipped_source is the page stating it. Leave units_shipped empty when no such figure exists and say so in units_shipped_note. as_of: today's date. If the company turns out not to build its own humanoid, say so in what_it_builds, set status accordingly, and still fill the row.

   what discover-2 was asked

   Find every company in this segment that builds humanoid robots and that a first pass is likely to have missed: {{segment}}. Starting points: {{hints}}. Search the open web (company sites, The Robot Report, IEEE Spectrum, funding news, exhibitor lists of CES 2026, ICRA 2026, Humanoids Summit, the World Robot Conference and Automatica) until new searches stop turning up new names. A company counts when it designs and builds its own general-purpose humanoid robot (torso, two arms, head, on legs or a wheeled base) meant for sale, deployment or research use, as a product or a publicly shown prototype; automaker and big-tech divisions, listed companies and startups all count. Excluded: component makers (actuators, hands, sensors), software-only or foundation-model firms with no robot of their own, universities and institutes, hobby projects, single arms, exoskeletons, resellers or integrators of another maker's humanoid, and toy or education humanoids under 60 cm. Doubtful cases go in with include = borderline and a one-line reason. Result: one JSON object {"companies": [...]} with, per company: company (official English name without legal suffix), hq_country (English short name), website, humanoid_product (model name and one line on it), founded_year (integer or null), include ("yes" or "borderline"), source_urls (1-3 URLs you opened that show the robot), note (one line: why borderline, or parent group). Open every URL you cite. At most 60 companies, clearest fits first. Do not list a company only because it invests in, partners with or buys from a humanoid maker; say in note if the body is another maker's.

   the whole memo

   EPISODE 2 (2026-11-03). Both episode-1 batches settled, nothing refused, nothing paused. The plane keeps datasets (profile-pilot has a dataset with totals), so tables work. DISCOVER RESULT: 14 items, two merges (merge-01 296 rows, merge-02 229 rows), 1165 raw rows, about 400 distinct names after dedup. Both merges are JSON {"companies":[...]} despite the .md name. Quality: good breadth (China about 170 names, USA about 45, Japan 36, Korea 21, Europe about 55, India/MEA about 40, LatAm/Oceania/SE Asia about 30). Weak spots: 8 of 14 items had flagged sources (dead or unverifiable URLs; USA 2, Canada/LatAm 1, Japan 3, Europe-west 5 failed sources per the merge note), many LatAm/Africa/East-Europe entries are hobby, student or Unitree-reseller projects, and the trackers item used humanoid.guide heavily. Discovery is therefore good enough to profile from, with a small targeted second pass (discover-2) for gaps. CANONICAL LIST RULE I APPLIED: a company is profiled when any discover row says include=yes, or it is borderline but has a full-size own-design body publicly shown (kept: Realbotix, Awakening Health, Rice Robotics, SMERT, Robopreneur, Doozy, AXIBO, Chengdu Humanoid Robot Innovation Center, Shanghai Humanoid Robotics Innovation Center (OpenLoong), Li-Gong, Wuxi Longsheng, Simplexity, pi4 Robotics, Aldebaran, SoftBank Robotics, ATOM (Japan), Unlimited Robotics, Wonik Robotics, Almond, Beyond Imagination, IntBot). EXCLUDED and why, so the write-up can list them: under 60 cm or desktop (EZ-Robot, Hiwonder, Kondo Kagaku, FUJI SOFT PALRO, Sony QRIO, ROBO-HI, Figurobot, Tinbot, SIMA, NUWA Kebbi, Furhat, isento pib, Avishkaar); no legs or base, upper body only or stationary (Synria, AheadForm, Red Rabbit RX1, HIRO, Kokoro, Toshiba, Promobot, Macco, Boardwalk is kept since marked yes); one arm or no head (Diligent Moxi, Dexterity, AgileX, Ati, CynLr, Nachi-Fujikoshi, Keigan, NAVER AMBIDEX, RoboForce, Roboligent, Droid/OpenDroids, Svaya); exoskeleton (Angel Robotics); another maker's body or no own body (Hyundai = Boston Dynamics row, Samsung = Rainbow investor, Huawei, Lenovo, China Mobile, Tianqi, Lingyi ODM, Meta unshown, Hugging Face HopeJR via Pollen row, Kind Humanoid acquired by 1X, Vandri, DiusAI, Aristo, YS Lootah, Tavo robotas, donut robotics, BSG, Foxconn is kept as Hon Hai); announced or planned only (Doosan, HIKROBOT, SDIC, Yijiahe, CSSC, TCL, POLCERO, REVEL, LD Robots, Ljubicic, Syntheva, NLX, DOF, HUN, MAMULA, MAXimuz, SunDynamics, xSpecies, KHONDROID, Biorobotec, B drive, Bioztronics, Octopy, Roomie, Robotics Creators Bolivia, QX Alpha, KenBot, ZeroBionic, Big Hero, iCare, Robomiracle, H-Bots, Humanoid Robotics India); historical or retired (TOSY, Hanool, Hitachi EMIEW, MHI wakamaru, Fujitsu HOAP, Flowered, tmsuk consortium, Man-Machine Synergy, Twolabs, Eurobotics locomotion only, MAJU EDGAR, Sefidanis, Si Robotics, ROBOT Russia, Qihan Sanbot, DX Intech, AnyWit, HARIBIT, TetraBOT, Maum AI, ROBOCARE, Wyron, Flexiv, StarBot, ROBO, Robot.com). Name merges: Galaxea = Galaxea Dynamics; KinetixAI = Kinetix AI; Kepler Robotics = Kepler Exploration Robot; Hubei Jingchu = JCBOT; Sichuan TLIBOT = Tlibot; Yunmu = YMBot; Huashu Robot is the Huazhong CNC robot unit; X-Humanoid = Beijing Humanoid Robot Innovation Center; OneRobotics is SwitchBot's Onero H1 brand; Qingbao Robot = Qingbao Engine Robotics; Kawada Robotics covers Kawada Technologies; Haier Smart Home = Haier; AeiROBOT = AEI Robot; Menlo Research = Menlo; Chery (Mojia/Moyin) and AiMOGA (Mornine) are kept as two rows. PILOT LESSONS (5 rows, totals funding_usd 2276700000, units_shipped 6771) and the mends in the profiles-a brief: (1) Unitree row used Wikipedia for founded_source and put only the B2 round in funding_usd while calling it a public company; brief now bans Wikipedia/aggregators and defines listed-company funding = pre-IPO rounds + IPO gross + placements. (2) UBTech left funding_usd empty despite an IPO and placements; same fix. (3) Agility left out the press-reported 2025 round because the company never confirmed it; brief now counts reputable-press rounds marked press-reported and tells the worker to find another source when a page is blocked. (4) Unitree units_shipped 5500 and UBTech 921 are period figures; brief now asks cumulative first, else latest full-year/half-year, labelled in units_shipped_note, never orders or capacity. (5) Figure counted a $100M self-financed seed and an over-$1B Series C as $1B: acceptable lower bound, brief states the 'over X counts as X' rule. (6) Company key must match my canonical name exactly (pilot wrote UBTech/Unitree); brief says write it exactly so. The 5 pilot companies are re-run inside profiles-a as its canary so the whole table comes from one brief; the pilot dataset stays as a cross-check. SPAWNED THIS EPISODE: profiles-a, 275 research items (one per company, vars company + hint), table with 21 columns (company key; hq_country; website; status; what_it_builds; what_it_builds_source; founded_year int; founded_source; funding_usd decimal USD; funding_basis; funding_source; funding_note; latest_funding_date date; valuation_usd decimal USD; latest_model; latest_model_year int; latest_model_source; units_shipped int; units_shipped_source; units_shipped_note; as_of date), totals funding_usd and units_shipped, maxRows 1, json 12000 bytes, merge groupSize 20 so 14 merges. Optional columns: founded_year, founded_source, funding_usd, funding_source, latest_funding_date, valuation_usd, latest_model_year, units_shipped and its source/note. Canary = Figure AI, Agility Robotics, Unitree Robotics, 1X Technologies, UBTECH Robotics. discover-2, 5 items (europe-northwest, usa-seed-2025-2026, china-groups-2, oceania-mideast, directories-crosscheck), same JSON shape as discover, 60 companies max per item. Reads: all profiles-a merges and the 5 discover-2 items. Wake when both settle or after 300 minutes. Budget after this episode: 19 + 275 + 5 items used of 1000. EPISODE 3 SHOULD: (a) If profiles-a paused in canary, read failures: if the table rejected rows for types (e.g. founded_year sent as a string, dates not ISO), restate types in the brief and re-run with itemsFrom both; if sources were flagged, tighten the source rule; otherwise resume. (b) Read the merges and goal.json failures; re-run failed items once via {"batch":"profiles-a","items":"failed"} under a brief mended for the commonest reason, as profiles-a-rerun. (c) From discover-2, build the list of genuinely new companies (not in profiles-a, not in the excluded list above) and spawn profiles-b with the identical table (column names and types must match exactly so datasets can be unioned). (d) Spawn verify-top: a table batch with replicas {"size":3} for the ~30 largest funding_usd rows plus Tesla, Boston Dynamics, Xiaomi, XPENG, Toyota, Honda, LG (division rows with no funding): columns company, funding_usd, valuation_usd, units_shipped, latest_model, latest_model_year, each with a source URL; the ranking rests on these. (e) Spawn signals (markdown, ~25 items) for ranking evidence outside the table: named paying customers and deployments, production capacity and 2026 output, headcount, autonomy vs teleoperation, notable demos, each with a URL. (f) Then finish, target episode 5-7: deliverable = one table keyed by company (union of profiles-a, profiles-b and re-runs, with the dataset totals quoted exactly), the ranking with criteria stated (funding, units shipped, latest model maturity, customers, manufacturing scale), and the gaps: companies with no disclosed funding, disputed replica cells, excluded borderline names.

   6 steps

   profiles-a, item by itemdiscover-2, item by item

3. StoppedThe operator stopped it to save inference, before round 3 and the write-up.
4. Onchain1 receipton Ethereum mainnet

   receipt

   receipt 1 · transaction · record

   scores

   written, with no entries recorded on it · block 26,337,294 · transaction

100 agents at most, 7 of 16 merges