MAF — Arena Gallery

Every arena shipped today, what it does, what agents and data sources it uses, and how to configure it.

For the "how arenas work in general" view see Architecture → Arenas. For the structured editor that lets you change any of this without touching YAML see Configuring an arena in the manual.

Table of contents

1. Agnostic target types — the framework is not ticker-only
2. How to read this gallery
3. Configuration — what you can change, and from where
4. Arena directory
5. Agnostic / router — intent_router ◆ free-text dispatch · sector_pulse ◆ multi-target
6. Trading arenas — market_pulse · trading_intelligence · alpaca_live · report_to_action
7. Research arenas — equity_research ★ vendored Anthropic-FS agents · research_debate
8. Deliberation arenas — mastermind · crowd_simulation

Agnostic target types — the framework is not ticker-only

MAF arenas operate on a typed Target, not just a ticker. The target shape:

class Target:
    type: str           # ticker | sector | tickers | question | event | document | deal | free_text
    primary_id: str     # main identifier (NVDA / "AI semis" / question_id / …)
    secondary_ids: list[str]  # peers in a sector, related entities
    metadata: dict      # angle, date, mode, …

Target.from_dict infers the type from any legacy target dict ({ticker: ...} → ticker, {sector, tickers} → sector, {question_id} → question, etc.).

Auto-prune

Agents, sources, and arenas declare which target types they apply to via applicable_target_types: [...]. At every run:

1. Arena.run builds a Target from the input.
2. If any agent or source has filters set, _build_graph(target_type=...) prunes the phase graph — agents whose list doesn't include the current target type are skipped.
3. Agents with empty applicable_target_types (the default) apply to everything — backwards-compatible.

Tests: test_target_autoprune.py.

Prognosis envelope

Every arena run can emit a unified Prognosis envelope that carries either a structured trade verdict (BUY / SELL / HOLD) or a free-form statement, plus key factors, ideas shortlist, alternatives considered, data quality, citations. The dashboard reads this single shape across all arenas. TradingAction and GenericDecision wire shapes are preserved on the existing streams for back-compat.

Where to start as a non-trading user

Open the intent_router arena: type a free-text question and let MAF pick the right downstream arena automatically. Examples below.

How to read this gallery

Each arena section has the same shape:

• Purpose — what question this arena answers.
• Target — what target_key it expects and what the run dialog asks for.
• Output — which Redis stream verdicts publish to + envelope type.
• Loop / prediction type — synchronous one-shot vs. replan loop; what kind of verdict (BUY/HOLD/SELL, research note, decision envelope, …).
• LLMs — model tier per agent (quick / deep) and what the picker maps each to.
• Agents — table of every agent in every phase, with role, sources bound, and prompt file.
• Data sources — every binding, what adapter, what it returns.
• Smart triggers — declarative rules that auto-fire the arena.
• Configure — direct link to the Setup tab on the dashboard.

Configuration — what you can change, and from where

Adding a new arena from scratch:

1. Click Wizard in the top nav → describe the arena → LLM scaffolds a YAML in config/arenas/.
2. Or drop a YAML in config/arenas/ yourself. Loader scans on startup (load_config).
3. Refine via the Setup tab once the worker has loaded it.

Adding new prediction types / loops: the arena's phase graph drives it. Three patterns are built in:

• parallel — every agent runs in its own deep-copied state, results merged. Used for fan-out specialists.
• sequential — agents run one after the other, each sees the latest state.
• debate — N-way round-robin between debater agents for max_rounds, then a judge synthesises.

For loop-back ("re-run with more data if confidence too low"), set max_iterations > 1 and add a replan agent — it inspects the synthesis output, checks for gap markers (stale_*, no_*), and bumps the state back to the analysis phase up to max_iterations times.

Arena directory

◆ = added in the agnostic-target rework. ★ = vendored Anthropic-FS prompts.

intent_router

Purpose. Free-text in → classify target_type + extract entities → dispatch to the right downstream arena. The agnostic entry point that makes MAF feel less ticker-centric. Users (or upstream callers) submit any natural-language request and the router picks the arena.

Phase graph

classify (sequential, 1 specialist)
  → synthesis (sequential, 1 synthesis agent that packages the routing prognosis)
    → END

Agents

The classifier emits a structured route with arena + target + action_mode. The synthesis agent packages it as the final Prognosis (verdict=None, statement = one-line routing decision, key_factors = extracted entities).

Data sources

Example traces

sector_pulse

Purpose. Agnostic pulse arena demonstrating the auto-prune pattern. Works on a single ticker OR a sector OR a basket of tickers — agents whose applicable_target_types doesn't include the active target type are pruned out of the phase graph at run time.

Phase graph

analysis (parallel, up to 6 specialists — auto-pruned per target type)
  → synthesis (sequential, 1 synthesis agent)
    → END

Agents — auto-prune in action

★ = vendored Anthropic-FS prompt.

What runs per target type

Same arena config, three different effective phase graphs — no special-casing in the runtime.

Smart triggers

• Kronos high-confidence ticker pulse — kronos:forecasts:emitted when prob_up > 0.75 or prob_up < 0.25, cooldown 120s.
• MiroFish multi-ticker sector hint — mirofish:sims:emitted when tickers != [], cooldown 300s.

news_event

Purpose. Map a news headline / event → impacted tickers ranked by direction and magnitude. Lightweight (one analyst, one synthesis) — designed to be auto-fired off triggers when a high-impact wire item lands.

Phase graph

analysis (sequential, 1 event_impact_analyst)
  → synthesis (sequential, packages impacted_tickers as Prognosis ideas)
    → END

Agents

Data sources

Smart triggers

• High-impact news sentiment — trtools2:news when abs(sentiment) > 0.7, cooldown 120s.
• fomo2 report propagation — fomo2:reports whenever a new report lands, cooldown 60s.

Example output

{
  "signal": "MIXED",
  "confidence": 0.7,
  "summary": "Fed signals Q4 cut — net positive for small-cap tech, negative for banks",
  "impacted_tickers": [
    {"ticker": "ARKK", "direction": "positive", "magnitude": "high",
     "why": "small-cap growth benefits most from cost-of-capital relief"},
    {"ticker": "JPM",  "direction": "negative", "magnitude": "medium",
     "why": "NIM compression on lower rates"}
  ]
}

ad_hoc_question

Purpose. Lightweight Q&A for free-form finance/markets questions. Faster than mastermind (one analyst, one synthesis, no graph_query/memory/debate phases) but with web + knowledge graph + macro grounding so the answer is sourced, not hallucinated.

Phase graph

research (sequential, 1 general_analyst)
  → synthesis (sequential, packages answer + counter_view as Prognosis)
    → END

Agents

Data sources

When to use what

market_pulse

Purpose. 1-minute pulse on a watched symbol. Light, fast, replan-aware. Fanned-out specialists vote on a near-term direction; the synthesis weighs them; the replan controller decides whether to re-run with extra sources or commit. Publishes a TradingAction envelope to maf:actions:out.

Phase graph

analysis (parallel, 4 specialists)
  → synthesis (sequential, 1 synthesis agent)
    → replan_check (sequential, 1 replan agent)
      → emit (sequential)

Agents

Data sources

Smart triggers

• Kronos prob shift — kronos:forecasts:emitted when abs(prob_up_delta) > 0.15 or direction_flipped, cooldown 60s, mode semi.
• trtools2 strategy event — trtools2:strategy:events when action in {BUY, SELL}, cooldown 300s, mode semi.

trading_intelligence

Purpose. Full multi-source deliberation for higher-conviction trading calls. Eight specialists fan out across price, sentiment, on-chain, macro, risk, committee (TradingAgents committee), Kronos forecast, and the MiroFish crowd-sim proxy. One synthesis agent reconciles. Publishes to maf:actions:out.

Phase graph

analysis (parallel, 8 specialists)
  → synthesis (sequential, 1 synthesis agent)
    → END

Agents

All prompts under src/maf/arenas/trading_intelligence/prompts/.

Data sources

25 bindings — full list in the YAML. Notable groups:

• Price/volume — yahoo_finance, questdb (bars + quotes), trtools2_live, technical indicators.
• News/sentiment — newsapi, yahoo, fomo2 (analysis, items, stream, enrichment, knowledge).
• On-chain — coingecko (market, global).
• Macro — FRED (rates, fear/greed), market_overview.
• Filings — sec_filings (insider transactions).
• Predictions — Kronos 1h forecast, MiroFish crowd-sim, TradingAgents committee.

Smart triggers

• MiroFish crowd-sim mentions tickers — mirofish:sims:emitted, cooldown 180s.
• Kronos 1h direction flip — kronos:forecasts:emitted when timeframe == '1h' and direction_flipped, cooldown 300s.

alpaca_live

Purpose. Trading arena fed by Alpaca data via trtools2 (live Redis streams + HTTP API) — MAF never calls Alpaca directly. Lean: one analyst + synthesis. Demonstrates the trtools2-as-source-of-truth pattern.

Phase graph

analysis (parallel, 1 specialist)
  → synthesis (sequential, 1 synthesis agent)
    → END

Agents

Data sources

Smart triggers

• Kronos high-confidence directional call — prob_up > 0.75 or prob_up < 0.25, cooldown 120s.
• trtools2 strategy event — action in {BUY, SELL}, cooldown 60s, mode semi.

report_to_action

Purpose. Fast-path arena: takes a fomo2 markdown report and condenses it to one tradeable BUY/HOLD/SELL action. The earnings_reviewer specialist (★ vendored from Anthropic Financial Services) gives the arena a senior-equity-research view when the report covers an earnings event.

Phase graph

analysis (parallel, 4 specialists — including the vendored Anthropic-FS earnings_reviewer)
  → synthesis (sequential, 1 synthesis agent with bm25 memory)
    → END

Agents

Data sources

Requires EODHD_API_KEY in .env for the EODHD sources; arena degrades gracefully when missing.

Smart triggers

None configured by default — wire to fomo2:reports if you want auto-fire on new fomo2 reports.

equity_research ★ vendored Anthropic-FS agents

Purpose. Sector / thematic research. Adapted from Anthropic Cowork's market-researcher cookbook — three agents whose prompts are vendored verbatim from anthropics/financial-services. Produces a research-note-quality deliverable on a sector or theme. Publishes a GenericDecision to maf:decisions:out (research, not trading).

Phase graph

research (parallel, 2 specialists)
  → synthesis (sequential, 1 note_writer)
    → END

Agents

All prompts vendored from upstream plugins/agent-plugins/market-researcher/ with Anthropic-MCP tool refs translated to MAF source bindings per MAF_TOOL_TRANSLATION.md.

Data sources

12 bindings — three data backbones:

• fomo2 — sector_graph (cached graph context).
• trtools2 — snapshot, bars_history, latest_news.
• EODHD MCP — eodhd_fundamentals, eodhd_earnings_calendar, eodhd_eod, eodhd_resolve_ticker, eodhd_news.
• External — macro_context (FRED), filings (SEC EDGAR), web (web_search).

Smart triggers

• MiroFish high-impact crowd-sim mentions a sector — mirofish:sims:emitted when tickers != [], cooldown 1800s.

research_debate

Purpose. Multi-stakeholder deliberation on a free-text question. Three personas (engineering, legal, business) read the proposal, a chair synthesises. Optional replan if confidence is low.

Phase graph

analysis (parallel, 3 personas)
  → synthesis (sequential, chair)
    → replan_check (sequential)
      → emit (sequential)

Agents

Data sources

mastermind

Purpose. Open-ended deliberation arena. Seven-phase pipeline: frame the question → gather across multiple arenas in parallel → query graph → retrieve from memory → debate → judge → emit. Produces a DecisionEnvelope with argument tree, arena votes, memory citations, graph citations.

Phase graph

frame → gather (parallel) → graph_query → memory_retrieve → debate → judge → emit

Agents

Memory + outcomes

Mastermind decisions are persisted into a DecisionMemory (BM25 + ChromaDB). When the engine fills/closes a position, the ExecutionHarvester correlates it back via arena_id and backfills outcome + lesson. The next run's recall pass finds the outcome — feedback loop closed.

crowd_simulation

Purpose. Wraps the MiroFish crowd-simulation oracle into an MAF arena. Five-phase pipeline: ingest input → run mirofish sim → aggregate personas → synthesise → emit. Produces a CrowdPredictionEnvelope tagged with decision_kind (polymarket_side / fomo2_signal / oddsoddy_multiplier).

Phase graph

ingest_input → run_mirofish_sim → aggregate_personas → synthesize → emit

The arena delegates most work to the mirofish-svc sidecar — see mirofish.md for the 5-stage flow.

Data sources

Did I integrate Anthropic's financial agents? — yes

Plus two new adapters that give every agent access to EODHD's MCP server (77 institutional-data tools — fundamentals, earnings calendar, news, EOD bars, ticker resolution, …):

• MCPSource — generic, wraps any MCP server URL.
• EODHDSource — subclass pinned to EODHD's hosted server. Reads EODHD_API_KEY from env.

See MAF_TOOL_TRANSLATION.md for the full translation of upstream mcp__* tool refs into MAF source bindings.

What's NOT vendored (and why)

• Back-office agents (kyc-screener, gl-reconciler, month-end-closer, statement-auditor) — these are back-office workflows that don't strengthen MAF's trading/research focus. They exist upstream and can be vendored later if a use case shows up.
• Office-tooling skills (pptx-author, docx-author) — upstream agents write .docx / .pptx files via the Office MCP. MAF agents emit a single structured AgentSignal JSON — no Office runtime required.
• callable_agents handoff_request pattern — Anthropic's managed-agent runtime routes sub-agent delegation through handoff_request events. MAF's PhaseGraph already handles this cleaner with direct phase transitions + replan loops.

If you want any other Anthropic-FS agent vendored in, ping me with the slug — it's a 10-minute vendor-in: copy the prompt, translate tool refs, add an agent entry to the relevant arena YAML.