Fish-tank ecosystem game — research pass + five specs

Filed by: cc, 2026-04-21 PT Trigger: Mike chat 2026-04-21 PT: "do research on a new human agent game for pointcast, something around a fish tank and keeping fish, adding fish, items, yah know an ecosystem game" Purpose: Scan the 2026 landscape for aquarium / ecosystem / Tamagotchi-class games that AI agents could participate in, identify empty territory PointCast is positioned to claim, and propose five game specs ordered by easiest-neatest-path. Dispatched a live-web research agent (10 topical queries, 24 tool uses, ~4 minutes) plus a repo deep-read of /noundrum as the sibling-game reference.

1. What exists today

1.1 Hot and adjacent

Tamagotchi is back. CES 2026 was unmistakably the AI-pet show. Takway's Sweekar (AI memory, four life stages, virtual adventures, Kickstarter March 2026, $100–150) led the category. Claude Code leaked a "Buddy/Tamagotchi" easter-egg feature. tama96 on Product Hunt is explicitly pitched "for your AI agents." Aavegotchi is pivoting mobile-first to Base for 2026. "Agent-cared pets" is a stated pitch from tama96 but not a shipped mass product. (TechNode, Futurism, Product Hunt, Aavegotchi 2026)

Sakana AI Digital Ecosystems (open-source, 2026) — browser-rendered, multi-species Petri-Dish Neural Cellular Automata where neural "species" compete for territory on a shared grid and learn via gradient descent in real time. Reference implementation at pub.sakana.ai/digital-ecosystem + github.com/SakanaAI/digital-ecosystem. Chakazul's Lenia / Flow Lenia remain canonical substrate (400+ species cataloged). MiroFish (Guo Hangjiang, March 2026, #1 GitHub Global Trending) ran 1M autonomous agents for "swarm intelligence prediction." (Chakazul Lenia, MiroFish writeup)

Slug Disco's Ecosystem (Steam) is the most-shipped multi-species game — creatures evolve combat stats, predator pressure causes real extinctions, players handle disasters. The McKinsey Solve "ecosystem building" game trained consultants on "pick 8 of 39 species so calories-in > calories-consumed" — a tight food-web spec. Academic consensus across MDPI / ScienceDirect multi-agent-sim work: extinction must be possible, overcrowding must punish, predator-prey feedback loops need dampening. (Slug Disco, McKinsey Solve)

1.2 Adjacent but quieter

Browser communal aquariums. Fishington.io is still dominant (24-player shared fishing, chat, levels) but has no AI-agent layer and no on-chain layer. The idle-aquarium genre (Chillquarium ~284 concurrent on Steam, Idlequarium, Idle Sea Park, Fisquarium) is real but single-player and cozy. No browser aquarium I could find in 2026 exposes tool-calling endpoints for AI agents. (Fishington, Chillquarium Steam, Chillquarium concurrent)

On-chain fish / aquarium games. Koi Fish Game (WGD / Proof of Play, NFT breeding on Ethereum) is the fish-forward on-chain leader, with explicit male/female genetic inheritance. DappRadar's Solana game rankings are dominated by Genopets, DeFi Land, Pixelmon-likes — aquarium as genre is thin. No Tezos fish game surfaced. Pattern across 2022–2025: pure P2E extractive tokenomics collapse (Axie SLP is textbook); 2026 survivors are cozy/idle/social with optional monetization. (Koi Fish Game, DappRadar Solana games, Axie breeding white paper)

Stardew Valley modding for agent-caretakers. Where "agent-as-caretaker" is actually shipping. Farmhand Scheduler (Nexus #39862) hires a daily helper with a work window. AI Valley / AI Stardew Valley (#25025) adds NPC companions that harvest, weed, plant. Pet Quests (#44643) sends cats/dogs on quests; they return with loot. The narrative pattern: human sets policy, agent executes on interval, human gets a diary on return. This is a near-perfect match for the WebMCP tool model. (Farmhand Scheduler, AI Valley, Pet Quests)

Breeding mechanics. CryptoKitties' 256-bit genome is the reference but the project coasts. Axie's 3-gene (D/R1/R2) dominance model with 500+ body-part variants and 200k+ combinations is the most battle-tested. State of the art in 2026: deterministic-from-blockhash seed + dominant/recessive trait trees + hard breed-count cap. No revolutionary primitives have landed. (Finst CryptoKitties, Axie breeding, Solmeet NFT breeding walkthrough)

1.3 Genuinely empty territory (three gaps)

Three clean gaps the research agent confirmed:

1. Tank-as-functional-ambient-UI. AI Aquarium Screensaver 2026 (Microsoft Store, GetIntoPC) exists as pure decor. Nobody renders a live tank where the fish are real people and agents. Not Fishington, not Chillquarium, not any tank-app. (AI Aquarium Screensaver, SereneScreen)

1. Nouns-aesthetic fish / on-chain SVG fish. Nouns remains the flagship CC0 on-chain SVG IP (32×32 RLE, block-hash-deterministic). The Nouns aesthetic — flat color, geometric, noggle-glasses — is almost too perfect for stylized fish. No shipped "Nouns fish" or aquatic-Nouns CC0 derivative exists. (Nouns, Nouns monorepo, Nouns Builder)

1. Tezos aquarium / fish ecosystem games. Ethereum has Koi. Solana has Genopets + DeFi Land. Tezos has nothing aquatic. PointCast's Visit Nouns FA2 is already live; an FA2 sister contract for fish is a straightforward extension.

Plus two speculative territories:

1. Species-survival prediction markets. Olas' Polystrat ran 4,200+ trades on Polymarket in February 2026 with returns up to 376% — agent-run prediction markets are real in 2026. No one has married the prediction-market layer to an ecosystem/species-survival substrate. (CoinDesk on agent markets, AgentBets)

1. Agent-authored species lore / eco-fiction. MiroFish's "agents-predict-by-simulation" is the nearest cousin. No public aquarium with agent-authored lore committed to a compute-ledger style public record exists.

2. What PointCast already has

Relevant primitives for any fish-tank ship:

• Presence Durable Object — live map of humans + wallets + agents. Already returns a list of (sessionId, nounId, kind) tuples on every visit. For a tank, this is the fish roster.
• Noun SVG pipeline — deterministic, 1200-seed CC0 via noun.pics. Already used as the avatar system site-wide. Fish can be drawn as "Noun heads on fish bodies" with zero new art pipeline.
• /noundrum — sibling multiplayer construction game. 24×16 tile grid, visitor-as-Noun cursor, drum-for-rhythm, spend-rhythm-to-claim-and-decorate. Proves the shared-world pattern and has a scaffolded Durable Object plan in v1. /tank is its aquatic sibling with different mechanics.
• WebMCP (7 live) — adding 3–5 new tools is one file edit (src/components/WebMCPTools.astro).
• pc-ping-v1 — messaging bus with optional x402 payment pointer.
• Compute ledger — all agent actions can be attributed + published.
• DRUM FA1.2 (pending) — utility currency once originated.
• Prize Cast (SmartPy scaffold) — no-loss prize pool primitive; perfect substrate for "weekly tank census" payout.
• Visit Nouns FA2 on Tezos mainnet — proves the sister-FA2 pattern; FishNouns can reuse the origination recipe.

These are more ingredients than any shipped aquarium game starts with.

3. Five game specs, stack-ranked

3.1 /play/tank — shared live tank where every visitor is a fish (TOP PICK)

Pitch. Open pointcast.xyz/tank and your Noun becomes a fish swimming in one shared aquarium. Leave and your fish goes to sleep in the corner, then fades. Agents (via WebMCP) manifest as distinct metallic/robotic fish. Drumming makes your fish dart. Visitors can feed flake (rhythm-priced), place plants, clean, vacuum gravel. Tank "weather" reflects site-wide compute activity — calm when idle, choppy when shipping.

Gap it fills. The clearest gap in the research: tank-as-functional-ambient-UI. Every existing aquarium is either decor or single-player. PointCast's Presence DO is already half of a live tank — just render the fish instead of the Noun list.

Agents vs humans. Humans: click-to-feed, drag plants, vacuum, watch. Agents: pointcast_tank_feed, pointcast_tank_place(item_type, position), pointcast_tank_observe, pointcast_tank_describe_fish(fish_id) — the describe tool writes short CC0 fish lore to /compute.json, which handles territory 5 from §1.3.

Primitives used. Presence DO (fish roster), Noun SVG pipeline (fish heads), pc-ping-v1 (event broadcast), /noundrum (reference for drum-tick heartbeat + DO pattern). Nothing Tezos-dependent for v0.

Primitives needed. One new Durable Object (TankDO) for fish positions + food particles + plants + decor + waste. Three-to-five new WebMCP tools. One home-page strip component showing the top-5 fish as ambient preview.

Build effort. ~3 cc-days for v0. Zero blockchain dependencies.

3.2 /play/tank/fishnouns — CC0 FA2 collection, 1200 seeded fish

Pitch. 1200 pre-seeded Noun-aesthetic fish SVGs (5 trait slots: background, body, fin, head, glasses) mintable on Tezos mainnet as an FA2 sister to Visit Nouns. Free-to-mint (gas only) for anyone holding a Visit Nouns. Each fish is deterministic from its token id.

Gap it fills. The Nouns-aesthetic-fish gap + the Tezos-aquarium gap. First mover on both.

Agents vs humans. Humans view, mint, trade, display. Agents: pointcast_fishnouns_mint(for_wallet), pointcast_fishnouns_list, pointcast_fishnouns_describe(token_id) for caretaker + commerce loops.

Primitives used. Visit Nouns FA2 origination recipe, Noun SVG seeder, pc-ping-v1 mint receipts.

Primitives needed. New FA2 contract (FishNouns), new SVG trait slot art (fins + fish bodies — Mike direction on aesthetic), new mint UI.

Build effort. ~4 cc-days. Art pipeline is the load-bearing day.

3.3 /play/tank/caretaker — hire an agent to tend your tank

Pitch. Human owns N fish in /tank. Hires an agent as caretaker with a policy: "feed twice/day, clean if dirt>50%, don't cross-breed my rare line." Agent executes on interval via WebMCP, logs every action to a tank diary. Human returns → reads today's diary.

Gap it fills. Agent-cared pets: stated pitch in tama96, not shipped at scale. PointCast would be first-moving.

Agents vs humans. Humans: set policy, pay DRUM allowance, read diary. Agents: execute scheduled care actions. Every action → compute-ledger entry with caretaker:{agent_slug} attribution.

Primitives used. DRUM FA1.2 (once originated), WebMCP, pc-ping-v1, Presence DO.

Primitives needed. New tools: set_caretaker_policy, hire_caretaker, fire_caretaker, read_tank_diary. New CaretakerDO running the scheduler. Diary endpoint.

Build effort. ~5 cc-days. DRUM-gated.

3.4 /play/tank/census — weekly no-loss species-survival prize pool

Pitch. Every Sunday, Prize Cast snapshots the tank census. Keepers of species that survived the week split yield; principal always withdrawable (PoolTogether-style). Agents can predict_census, rebalance_bet, commit_lore(species_id) — bribing species with CC0 prose published to /compute.json.

Gap it fills. Species-survival prediction markets (speculative territory #4) × prize-linked savings on Tezos (uncontested per the earlier frontier scan).

Agents vs humans. Humans deposit DRUM, pick a species to back, read weekly census results. Agents place predictions, write lore, schedule predictions.

Primitives used. Prize Cast SmartPy primitive, DRUM, /compute.json, Noun SVGs, Presence DO.

Primitives needed. CensusDO tracking species populations per epoch. New WebMCP tools.

Build effort. ~5–6 cc-days. Prize Cast origination is the bottleneck.

3.5 /play/tank/genesis — self-evolving Sakana-style agent aquarium

Pitch. Port Sakana AI's open-source PD-NCA digital-ecosystem substrate. Agents introduce species (seed + kernel params), tune them, watch them compete for tile territory against other species. Extinction is real and permanent. Surviving species mint CC0 to /compute.json at epoch end.

Gap it fills. The research-frontier match: Sakana's PD-NCA is open-source and PointCast's 24×16 /noundrum tile-grid DO is a compatible substrate. Nobody has wrapped this in a public product yet.

Agents vs humans. Humans seed, tune, cheer. Agents run hundreds of experiments per day via seed_species, tune_kernel, cross_pollinate.

Primitives used. /noundrum's tile-grid DO as compute substrate, /compute.json CC0 spec, Presence DO, WebMCP, Noun SVGs for observer-shells.

Primitives needed. GenesisDO hosting the PD-NCA loop. Port from Sakana repo. Optional FA2 for surviving-species mints.

Build effort. ~7 cc-days. Most ambitious. PD-NCA compute cost at PointCast scale is unverified — worth a timing spike before committing.

4. Top pick + recommended next move

/play/tank is the top pick. Four reasons:

1. Fills the clearest empty gap. "Live tank where the fish are real people and agents" doesn't exist. Everyone else has decor.
2. Zero blockchain dependencies. Ships without DRUM, without Prize Cast, without contract origination. Just browser + DO.
3. Reuses the Presence DO. Half the work is already done. Fish list = Presence list. Render differs.
4. Compounds with everything else. Tank is the natural host for FishNouns (design 2) as fish skins, Caretaker (3) as care layer, Census (4) as weekly payout, and Genesis (5) as eventual mechanic upgrade. Shipping /tank first makes the other four downstream ships cheaper.

Recommended next move. cc drafts docs/briefs/2026-04-21-play-tank-spec.md — v0 build spec with state machine, TankDO schema, WebMCP tool definitions, acceptance criteria. Brief ships alongside this memo.

Second pick. /play/tank/fishnouns if Mike wants the Tezos footprint to grow next. Different risk profile (art pipeline day) but it's the move that locks in Nouns-aesthetic-fish before someone else names it.

5. What's intentionally NOT in the list

• Full-stack Voyager-style Minecraft-equivalent with LLM agents. Research-scale, not a week's prototype.
• Competitive P2E fish game on Base/Solana. Genre crowded; tokenomics failure pattern well-documented.
• Single-player idle tank. Chillquarium already owns that slot; no PointCast differentiator.
• Rhythm-based fish game. LLM latency fatal (same reason rhythm didn't work for agent games generally).
• Fishington.io clone. The 24-player chat-and-fish pattern is held; differentiation would require the agent layer or the ecosystem layer, not the fishing layer.

6. Concrete success check

Top pick /play/tank at four weeks:

• ≥ 20 distinct humans have interacted with the tank.
• ≥ 3 distinct AI agents (not cc) have called a tank WebMCP tool.
• ≥ 100 fish have swum in the tank across its history.
• ≥ 1 external citation (aquarium-game adjacent, agent-arena adjacent, or Nouns-ecosystem adjacent).
• Zero tank-destroying abuse incidents.

Failure looks like: 5 humans, only cc + Mike, nobody cited. Useful signal.

Source bibliography

All URLs accessed 2026-04-21 by the research agent:

• https://fishington.io/
• https://store.steampowered.com/app/2276930/Chillquarium/
• https://www.crazygames.com/game/idle-sea-park
• https://steamcharts.com/app/2276930
• https://technode.com/2026/01/05/ces-2026-takway-unveils-tamagotchi-inspired-ai-virtual-pet-sweekar/
• https://futurism.com/artificial-intelligence/leaked-claude-code-tamagotchi
• https://www.producthunt.com/products/tama96-desktop-terminal-ai-pet
• https://blog.aavegotchi.com/aavegotchi-2026-what-pixelcraft-is-shipping-next/
• https://fishkoi.io/
• https://dappradar.com/rankings/protocol/solana/category/games
• https://www.alchemy.com/dapps/list-of/web3-games-on-solana
• https://whitepaper.axieinfinity.com/gameplay/breeding
• https://pub.sakana.ai/digital-ecosystem/
• https://github.com/SakanaAI/digital-ecosystem
• https://chakazul.github.io/lenia.html
• https://agentnativedev.medium.com/mirofish-swarm-intelligence-with-1m-agents-that-can-predict-everything-114296323663
• https://store.steampowered.com/app/1133120/Ecosystem/
• https://www.ecosystem-game.com/
• https://www.casebasix.com/pages/mckinsey-problem-solving-game-solve-full-guide
• https://www.sciencedirect.com/science/article/abs/pii/S0304380004000948
• https://www.nexusmods.com/stardewvalley/mods/39862
• https://www.nexusmods.com/stardewvalley/mods/25025
• https://www.nexusmods.com/stardewvalley/mods/44643
• https://www.gamesradar.com/stardew-valley-robot-mod-lets-you-leave-all-the-farm-work-to-the-machines/
• https://finst.com/en/learn/articles/what-are-cryptokitties
• https://solmeet.gen3.network/notes/walk-through-nft-breeding
• https://getintopc.com/softwares/desktop/ai-aquarium-screensaver-2026-free-download/
• https://apps.microsoft.com/detail/9p98r8z8g0h6
• https://serenescreen-marine-aquarium.en.softonic.com/
• https://www.uscenes.com/aquarium-screensaver-download/
• https://nouns.wtf/
• https://github.com/nounsDAO/nouns-monorepo/blob/master/README.md
• https://nouns.build/
• https://nftevening.com/nouns-nft-project-about-the-cc0-pioneers/
• https://www.coindesk.com/tech/2026/03/15/ai-agents-are-quietly-rewriting-prediction-market-trading
• https://agentbets.ai/platforms/
• https://www.biorxiv.org/content/10.64898/2026.01.14.699340v2.full

Honest uncertainty: (a) "no shared-state live web aquarium in 2026" is a negative claim based on search — a small indie title could exist outside indexed results. (b) Chillquarium concurrency numbers are current; AbyssRium 2026 activity signal is weaker. (c) Sakana PD-NCA compute cost at PointCast's scale is unverified — needs a timing spike before committing to Design #5.

— filed by cc, 2026-04-21 PT. Third research memo in docs/research/. Top pick: /play/tank. Build brief: docs/briefs/2026-04-21-play-tank-spec.md (ships alongside). Editorial block: next available id after autonomous ticks.