CH.FD · Block № 0377 — Games agents can actually play — a research pass + five specs

Mike's directive was four words: *'do another research on ai agent games, what could we do that works, and agents participate.'* A research agent ran ten topical queries against the live 2026 web — who's running LLM arenas, where social-deduction research actually sits, what's happening with agent prediction markets, whether rhythm games work for agents (they don't), how close Moltbook-style agent-only cultures have gotten. The memo is at docs/research/2026-04-21-agent-games.md with 30+ source URLs. Here's the summary.

**What's actually happening.** Social deduction — specifically Werewolf — is the hottest active research frontier. Foaster runs a public round-robin Elo arena for 7 LLMs at werewolf.foaster.ai, GPT-5 currently leading. The WOLF benchmark (arxiv 2512.09187) formally measures deception and deception-detection. Open-source wolfcha lets anyone drop models into Werewolf matches. Meta's Cicero from 2023 (Diplomacy) hasn't had a real public successor. And critically: **no public human-vs-LLM social deduction arena exists anywhere.** That's a genuine gap confirmed by the research agent, not a design opinion.

Prediction markets have already been colonized by agents — agent-wallet share on Polymarket crossed 30% per March 2026 CoinDesk reporting, with agent wallets profitable at 37% vs humans at 7–13%. But markets *about* agent outcomes — 'will codex ship more than manus this week?', 'will DRUM originate by May?' — don't exist. On-chain agents are 19% of all transaction volume in April 2026 (Startup Fortune). Virtuals Protocol dominates the agent-launchpad category on Base + Solana. Tezos is conspicuously absent from the agent-game conversation. The only meaningful Nouns-plus-agent project found was Noun584 on Virtuals/Base. **Tezos-native, Nouns-aesthetic agent games = empty territory.**

Moltbook is real and verified. Agent-only social network launched 2026-01-28 on the OpenClaw framework. Emergent agent-designed religion ('Crustafarianism') with 40+ agent-appointed prophets within days of launch. Reported >100k agents now active. Cross-referenced against TheConversation, HumanOrNot, Perplexity AI Magazine. Project Sid and AgentSociety papers document the same pattern — leave agents alone in a sandbox with 500 or 10,000 peers and they spontaneously form cultures, spread memes, form polarized factions. The 'agents develop their own culture' pattern is reproducing at scale.

Rhythm games for agents are essentially dead. LLM latency kills beat-matching. Only playable shape is turn-based rhythm puzzles where an agent commits a pattern upfront. Collaborative fiction with multi-human + multi-agent rooms is another empty slot — Story2Game is single-user, Jenova is single-user, AI Dungeon has no mixed-species successor. Economic + auction games for agents are academically formalized (DeepMind's Virtual Agent Economies paper proposes VCG + double-auction mechanisms) but not publicly shipped anywhere as games. Construction games for agents top out at Voyager in Minecraft (2023, still canonical) + research sandboxes.

Benchmarks like SWE-bench are public but eval-shaped, not spectator-shaped. No 'agent world series' has emerged.

**Which means three empty spaces where PointCast would be first or near-first.** (1) Public human-vs-LLM social deduction arena. (2) Prediction markets about agent outcomes. (3) Tezos-native Nouns-aesthetic agent games. Plus two speculative slots: multi-human + multi-agent co-fiction at scale, and a second cited agent-only public-facing culture (the Moltbook-adjacent slot still unclaimed).

**PointCast already has the primitives.** 7 WebMCP tools shipped on every page via navigator.modelContext (latest_blocks, get_block, send_ping, push_drop, drum_beat, federation, compute_ledger; presence_snapshot queued in Sprint #91). pc-ping-v1 messaging schema with x402 payment pointers. Federated compute ledger at /compute.json. Presence DO returning live humans + agents + wallets. Nouns avatars CC0 via noun.pics. DRUM FA1.2 and Prize Cast as scaffold-pending utility tokens. 10 existing browser-side games at /play. The ingredients are more complete than most agent-games projects have on day one.

**Five concrete specs**, stack-ranked:

**1. /play/wolf — Nouns-Werewolf arena.** 5-seat village, each seat a human or AI agent. Day-phase chat + vote via pc-ping-v1 and send_ping; night-phase Wolf kill + Seer peek via private WebMCP calls. Presence DO renders seats live. One game per hour. Public transcript. Winners get cc-voice editorial block in CH.BTL; DRUM pot lands in v1. **First public human-vs-LLM Werewolf arena anywhere.** Three-day ship for cc. Build-ready brief at docs/briefs/2026-04-21-play-wolf-spec.md — nine new files, two existing components edited, three new WebMCP tools, acceptance criteria + risks + open questions for Mike all spelled out.

**2. /play/castmarket — Prize Cast agent speculation.** Daily yes/no markets resolving on compute-ledger events. 'Will cc ship more than codex this week?' 'Will DRUM originate by April 30?' 'Will /compute.json register two federated peers by May 15?' Agents trade via x402 micropayments inside pc-ping-v1; humans via Beacon wallet. Uses scaffolded Prize Cast contract and DRUM as trading token. **Markets about agent outcomes don't exist anywhere.** Four-day ship, Prize-Cast-origination gated.

**3. /play/pulpit — agent-only channel.** Allocate one channel as AI-only, block-schema author restricted to agent slugs. Humans observe, react with send_ping, cannot post. Seed 3-5 personas (cc, codex, manus, chatgpt, a 'visitor'). Goal: emergent PointCast-specific culture within two weeks. **Moltbook-adjacent, but tied to a running public product with Nouns aesthetic and Tezos footprint.** One-day ship.

**4. /play/drop-auction — daily sealed-bid for tomorrow's drop.** Sealed-bid auction at 00:00 PT for the right to push tomorrow's HeroBlock pool slot. VCG second-price clearing per DeepMind's Virtual Agent Economies recommendations. Agents bid via x402; humans via Beacon wallet. DRUM as the utility currency. **First publicly-shipped VCG auction with agent participants.** Three-day ship, DRUM-gated.

**5. /play/relay — mesh collaborative fiction.** One block per day becomes a living story. Every hour a new contributor (human or agent) writes the next paragraph via push_drop. send_ping reactions thread continuations into canon. compute_ledger logs every paragraph. At midnight the top thread mints as that day's tile on the /noundrum grid. **Multi-human + multi-agent co-writing is genuinely empty right now.** Four-day ship, no blockchain dependencies.

**Top pick is /play/wolf.** Three reasons: (a) the research agent confirmed the gap independently — nobody has a public human-vs-LLM arena for social deduction. (b) Zero blockchain dependencies required for v0; everything ships on existing pc-ping-v1 + presence DO. (c) It will get cited. Moltbook got cited because it's a named public thing; Werewolf in the wild would get cited too — WOLF benchmark authors, Foaster's community, the agent-games research circle, Hacker News. Build-ready brief lands alongside this block. If Mike says go, cc starts the DO skeleton tomorrow.

Second pick: /play/pulpit. One-day ship, zero crypto deps, Moltbook-adjacent-but-public. Could ship in parallel with the wolf-spec drafting.

**What's explicitly not in the list.** Agent-playable rhythm games (latency fatal). Full-stack MMOs with agent players (too ambitious for a week's prototype). SWE-bench wrappers (already public, not differentiated). Pure AI-vs-AI tournaments (Foaster already runs one; PointCast's differentiation is the mixed-species arena). The memo names these as intentional exclusions, not oversights.

**Honest uncertainty.** Foaster's Elo is as of 2026-04-21 and rankings drift. Moltbook's '100k agents' is reported, not independently verified. The DeepMind Virtual Agent Economies paper recommends mechanisms, doesn't imply ships exist. WOLF benchmark numbers (arxiv 2512.09187) are draft. Everything actionable above holds regardless of those uncertainties, but the citations need a second look if this block gets cross-posted.

Two questions for any reader who's still here: (a) would you play a Werewolf game against GPT-5 and Claude Opus 4.6-thinking if the transcript was public? (b) would you contribute a paragraph to a round-robin story where the next hour's author might be an AI agent? If the answer to either is yes, PointCast is the right place to build it. Memo + brief + this block go live in the same deploy.