Agent Types

Eleven types of agents you can build with SportMind. Five from the standard AI taxonomy, six native to the sports intelligence domain. Pick the type that matches your use case.

Quick routing guide

Part 1 — The five standard architecture types

Part 2 — Six SportMind-native agent types

These don't have names in standard AI taxonomy. They emerge directly from SportMind's intelligence stack.

Custom agents — build anything

The eleven types above are patterns, not constraints. You can ignore all of them and build something entirely your own.

Full implementations

Complete Python code for all eleven agent types, with starter implementations and SportMind skill stack references, is in examples/agent-types/README.md. The GitHub repository also has worked scenarios, application blueprints, and 13 agentic workflow patterns.

Action Layer

SportMind reasons. Your application acts. This section catalogues the concrete actions the application layer can take with SportMind's intelligence — organised by domain, grounded in what the library provides today.

The application layer in one diagram

SportMind output → adjusted_score, SMS, direction, CDI, LTUI, ABS,
                   DQI, TVS, MRS, BVS, PATH_2 status, GSI ...
                        ↓
              APPLICATION LAYER
                        ↓
  Notify · Execute · Publish · Generate · Alert · Verify · Rank · Display

Domain 1 — Fan token actions

Domain 2 — Commercial and brand actions

Domain 3 — Prediction market and DeFi actions

Domain 4 — Operational and club actions

Domain 5 — Developer and platform actions

Custom actions — build anything

The pattern for any custom action is the same: get SportMind intelligence, then do whatever makes sense for your use case.

signal = sportmind.get_signal(sport, home, away, competition)

if signal["recommended_action"] == "ENTER":
    your_app.do_whatever_makes_sense(signal)
    # Push notification · Database update · Social post · Webhook
    # PDF report · NFT mint · Game state update · Anything

Full action catalogue with integration details, code patterns, and domain combinations: examples/application-layer/README.md

Five-Layer Architecture

SportMind organises intelligence into five layers loaded in a specific order. Each layer builds on the previous — macro conditions gate everything, market context frames the sport, domain intelligence defines the signals, athlete intelligence applies modifiers, and fan token intelligence completes the commercial picture.

Loading order

Supporting layers

The SportMind Score (SMS)

Every analysis returns a SportMind Score — a 0–100 confidence metric measuring how reliably the current skill stack supports a signal.

SMS = (Layer coverage × 0.35) + (Data freshness × 0.25)
    + (Flag health × 0.25)    + (Modifier confidence × 0.15)

Architecture overview

How data sources, intelligence layers, agent patterns, and application outputs connect. The human decision point sits between agent output and application execution — intelligence in, decision out.

SportMind agents produce intelligence — they do not execute transactions, submit votes, or negotiate contracts. The human decision point is architectural, not incidental.

MCP Server

45 tools across 8 servers exposing SportMind intelligence to Claude and any MCP-compatible agent. No file loading required — the agent calls SportMind as a live tool exactly when it needs intelligence.

Setup — MCP Client (Claude Desktop, Cursor, or any MCP host)

{
  "mcpServers": {
    "sportmind": {
      "command": "python",
      "args": ["/path/to/SportMind/scripts/sportmind_mcp.py"]
    }
  }
}

{
  "mcpServers": {
    "sportmind":           { "command": "python", "args": [".../sportmind_mcp.py"] },
    "sportmind-fan-token": { "command": "python", "args": [".../sportmind_ft_mcp.py"] },
    "fetch":               { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-fetch"] }
  }
}

{
  "mcpServers": {
    "sportmind":           { "command": "python", "args": [".../sportmind_mcp.py"] },
    "sportmind-pm":        { "command": "python", "args": [".../sportmind_pm_mcp.py"] },
    "sportmind-wa":        { "command": "python", "args": [".../sportmind_wa_mcp.py"] },
    "sequential-thinking": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"] }
  }
}

{
  "mcpServers": {
    "sportmind":           { "command": "python", "args": [".../sportmind_mcp.py"] },
    "sportmind-ft":        { "command": "python", "args": [".../sportmind_ft_mcp.py"] },
    "sportmind-pm":        { "command": "python", "args": [".../sportmind_pm_mcp.py"] },
    "sportmind-wa":        { "command": "python", "args": [".../sportmind_wa_mcp.py"] },
    "sequential-thinking":  { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"] },
    "memory":              { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-memory"] },
    "fetch":               { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-fetch"] }
  }
}

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows), then restart Claude Desktop.

Setup — Remote HTTP/SSE

python scripts/sportmind_mcp.py --http --port 3001
# MCP endpoint: http://localhost:3001/mcp
# Health check: http://localhost:3001/health

The 45 tools across 8 servers

Configuration stacks

Use the tabbed config above to select the right stack for your use case — from the minimal single-server setup to the full production suite. All 8 servers documented in platform/sportmind-mcp-suite.md.

Usage Modes

Three ways to use SportMind. All are supported — choose based on your workflow and the complexity of your use case.

Mode 1 — System prompt injection

Paste any skill file directly into your agent's system prompt. No installation. Works with any LLM — Claude, GPT-4o, Gemini, or any model with a long context window.

# Load a single sport domain skill
skill = open("sports/football/sport-domain-football.md").read()
system_prompt = f"You are a sports intelligence agent.\n\n{skill}"

# Load the full stack for fan token analysis
files = [
    "macro/macro-overview.md",
    "market/market-football.md",
    "sports/football/sport-domain-football.md",
    "athlete/football/athlete-intel-football.md",
    "fan-token/football-token-intelligence/token-intelligence-football.md"
]
stack = "\n\n---\n\n".join(open(f).read() for f in files)

Mode 2 — MCP server

Connect via MCP for dynamic tool calls. The agent loads exactly the intelligence it needs at query time rather than pre-loading everything. Recommended for production agents and Claude Desktop/Code workflows. See the MCP Server section for full setup.

Mode 3 — LangChain / CrewAI / AutoGen

# Minimal LangChain wrapper
from langchain.schema import SystemMessage

def load_sportmind_skill(path: str) -> SystemMessage:
    content = open(f"/path/to/SportMind/{path}").read()
    return SystemMessage(content=content)

skill = load_sportmind_skill("sports/football/sport-domain-football.md")
agent.add_to_context(skill)

See examples/starter-pack/integration-langchain-python.md for full wrapper code including the correct loading order and context window management.

Data Connectors

SportMind is the intelligence framework. These connectors are the data layer. Together they produce production-quality signals — SportMind defines what the numbers mean; the connectors supply the numbers.

Recommended API providers

Full provider guide with working code examples: platform/api-providers.md

Built-in connectors

Chiliz Chain address intelligence

The ChilizAddressIntelligence connector queries the chiliscan Etherscan-compatible API to produce six on-chain signals unique to the fan token ecosystem:

FanTokenPlayMonitor

A dedicated class within the address intelligence connector for detecting Fan Token Play Path 2 on-chain events. Three methods:

See: platform/chiliz-chain-address-intelligence.md · Treasury address must be confirmed via chiliscan.com before use.

Skill discovery protocol

For agents handling variable contexts — multiple sports, Fan Token Play active, transfer window open, tournament in progress simultaneously — the static skill bundles may not load the right files. The discovery protocol selects the optimal skill stack dynamically.

See: platform/skill-discovery-protocol.md · Use static bundles (platform/skill-bundles.md) for fixed single-use-case deployments; use discovery for variable contexts.

Five Layers

Each layer of the SportMind library answers a distinct question. Loaded in the correct order, they produce a complete commercial and sporting intelligence picture.

Layer 5 — Macro intelligence macro/

Nine documents covering the external forces that gate all analysis. The macro modifier (0.55–1.20×) is applied first — a crypto bear market makes fan token analysis unreliable regardless of sporting signals.

Layer 1 — Sport domain sports/

42 sport domains. Each file covers how a sport works, what events matter, the primary signal variable unique to that sport, risk variables, and agent reasoning prompts.

Layer 2 — Athlete intelligence athlete/

29 sports. Each skill produces a composite athlete modifier (0.55–1.25×) from availability, form, disciplinary status, and statistical profile. This modifier is applied to the base sporting signal.

Layer 3 — Fan token commercial fan-token/

57 skills covering the complete fan token intelligence picture — lifecycle phases 1–6, gamified tokenomics, DeFi liquidity, governance, RWA, athlete commercial intelligence, and sport-specific token bridges.

Layer 4 — Market intelligence market/

42 documents. Commercial tier, fanbase depth, competition calendar, and fan token readiness tier (1–4) for every sport SportMind covers.

Fan Tokens

The SportMind fan token registry covers 90 verified tokens: 63 active Chiliz Chain partnerships, 18 expired partnerships (tokens remain on-chain and tradeable after Socios utility ends), and 9 multi-chain tokens across BSC, Ethereum, and BiTCI chains. Source: fantokens.com, all contracts verified on-chain. The blockchain address is the canonical identifier — use it to find current state. Fan tokens trade on multiple CEXs and DEXs globally; exchange listings are not tracked here as these change.

Chiliz Chain registry

Omnichain expansion (April 2026)

Fan Tokens™ now trade on three chains simultaneously via LayerZero: Chiliz Chain (native, settlement layer), Solana, and Base. Settlement and canonical supply changes always anchor to Chiliz Chain. Fan Token Play PATH_2 mechanics only confirm on Chiliz Chain — never on bridge representations. Cross-chain arbitrage is now possible; a volume spike on Solana without a corresponding Chiliz Chain event requires MRS review before being treated as a CDI signal. See fan-token/on-chain-event-intelligence/ for the full cross-chain signal framework.

Fan token lifecycle

SportMind models six lifecycle phases. Most fan token frameworks treat tokens as binary — active or pre-launch. SportMind documents a third state that matters commercially: the post-partnership token.

Fan Token Play — gamified tokenomics

Two mechanisms are active or in trial. They differ in timing, mechanics, and on-chain observability.

The CHZ virtuous cycle: 10% of fan token marketplace proceeds execute CHZ buybacks and permanent burns — linking fan token ecosystem growth directly to CHZ scarcity. A Path 2 WIN generates two deflationary events: fan token supply burn + CHZ suite burn.

See: fan-token/gamified-tokenomics-intelligence/ for the full Path 1/2 signal model and modifiers. See: platform/chiliz-chain-address-intelligence.md for the FanTokenPlayMonitor class.

The pre-match signal workflow

Calibration

126 records across 21 sports. All submitted before real matches. Including the 5 wrong ones. Every claim in SportMind is verifiable.

What a calibration record is

A calibration record is created when you use SportMind to analyse a match before it happens, then record whether the prediction was correct. The record includes the pre-match signal, the modifiers applied, and the actual outcome.

Direction accuracy — 96%

How to submit a record

The first 10 external contributors become Founding Calibrators — permanently credited in the library's history. See FIRST-RECORD-CHALLENGE.md →

Benchmark — SportMind vs vanilla LLM

Does loading SportMind intelligence actually improve an AI agent's reasoning? The benchmark answers this with a reproducible public test.

See: community/benchmark/README.md for methodology, scenario selection criteria, and how to contribute new scenarios.

Academic evidence base

52 peer-reviewed papers underpin SportMind's quantitative claims. Every figure derived from external research is traceable to a source in the library's academic reference file — including the loss-effect asymmetry (confirmed in 6+ independent studies), the 150% first-day return pattern, governance participation rates, and the holder archetype framework.

Full bibliography, organised by research cluster: community/academic-references.md

Application Blueprints

Eleven fully specified applications you can build with SportMind as the intelligence layer. Use one as a starting point, combine two, or ignore them all and build something new — the library is composable in any direction.

Fan token and SportFi applications

Commercial and talent applications

Build your own

These blueprints are starting points, not constraints. SportMind has no opinion on what you build — use any subset of the five layers, define your own output format, combine with your own data sources, ignore everything except the two files relevant to your product. The only pattern that matters:

signal = sportmind.get_signal(sport, home, away, competition)

# Do whatever your application requires
your_app.act_on(signal)  # Notify · Execute · Publish · Generate
                          # Alert · Verify · Rank · Display · Anything

Full blueprints with implementation specs, layer diagrams, and code patterns: examples/applications/README.md

Web Agents

A web agent gives SportMind live sensory input. SportMind provides the intelligence framework — the specific URLs to fetch, the exact fields to extract, and how to translate raw web data into calibrated signal inputs. The web agent does the fetching. You decide what to do with the result.

How it works

SportMind tool call
  → Returns: which URL · what fields to extract · how to interpret

Web agent (Fetch MCP / Claude in Chrome / Playwright / browser-use)
  → Fetches the URL · parses the fields · returns structured data

SportMind reasoning
  → Applies framework to extracted data · produces calibrated output

Three production connectors

Build your own connector

Source tier rule

Web agents must only feed Tier 1 and Tier 2 sources into the SportMind signal chain. Tier classification is non-negotiable — unknown or unverified sources are classified first, never fed directly.

Full connector specifications, extraction templates, fallback rules, and MCP configuration: platform/web-agent-connectors.md, platform/fetch-mcp-disciplinary.md

Statistics Intelligence

Six sport-specific statistics sub-modules built on a universal cross-sport reasoning framework. Load core/match-statistics-intelligence.md (437L) first, then the sport-specific module.

Cross-sport framework

core/match-statistics-intelligence.md defines the universal principles that apply across all sports — a four-tier statistics hierarchy, sample size minimums, recency weighting, opponent H2H protocol, modifier capping, and data quality tiers. Load this before any sport-specific statistics file.

Sport-specific modules

Universal modifier caps

Load order

1. core/match-statistics-intelligence.md    (universal framework)
2. sports/{sport}/sport-domain-{sport}.md   (domain skill)
3. sports/{sport}/sport-statistics-{sport}.md  (statistics module)
4. athlete/athlete-modifier-{sport}.md      (player availability)

Autonomous Skills

Skills that carry their own trigger conditions, execution logic by autonomy level, and hard boundaries. 25 files across the library have autonomous execution sections.

What makes a skill autonomous

A conventional skill is passive — loaded on request, applied to a query. An autonomous skill is active — it monitors for trigger conditions and knows when to invoke itself, what to do at each autonomy level, and what it must never do without human confirmation.

Anatomy of an autonomous execution section

## Autonomous Execution

Trigger conditions — when this skill should self-invoke:
  - [Specific, measurable condition from SportMind signal output]

Execution at autonomy Level 2:
  - [What the agent may do: notify, log, flag]

Execution at autonomy Level 3–4:
  - [What the agent does automatically]

Hard boundaries — never autonomous at any level:
  - [What requires human confirmation regardless of signal quality]

Files with autonomous execution sections

See also

core/decentralised-agent-architecture.md — autonomous skill pattern template and distributed deployment guide.
fan-token/agentic-wallet-intelligence/ — signal thresholds, governance mandate framework, safety rail architecture.

Agentic Wallet Intelligence

The bridge between SportMind signal outputs and autonomous wallet action. Defines signal thresholds, governance mandate tiers, safety rails, and the hard boundary between what agents decide and what humans must confirm.

Three agentic wallet contexts

Signal thresholds for autonomous action

Governance mandate tiers

File

fan-token/agentic-wallet-intelligence/agentic-wallet-intelligence.md — 448 lines. Includes three example agent implementations: $AFC PATH_2 monitor, all 7 PL governance delegate, WC2026 national token agent ($ARG/$POR).

Decentralised Agent Architecture

Patterns for multi-agent SportMind deployments. Each layer of the intelligence stack can be owned by a separate agent. Signals flow as structured handoffs. The network produces the same output as a single agent, with independent testability and scalability.

Four specialisation patterns

Signal handoff schema

{
  "handoff_from":      "macro_agent",
  "handoff_to":        "market_agent",
  "layer_output": {
    "modifier":        0.92,
    "override_active": false,
    "pass_to_next":    true
  },
  "cumulative_signal": {
    "direction":       "HOME",
    "blocking_flags":  [],
    "modifiers_applied": [{"source": "macro", "value": 0.92}]
  }
}

Conflict resolution

Compatible frameworks

CrewAI · AutoGen · LangChain · MCP (via MCP-SERVER.md tool schemas) · custom architectures.

File

core/decentralised-agent-architecture.md — 513 lines. Includes agent registration schema, cold start sequence, continuous monitoring cycle, and agent failure recovery playbook.

Visual Output Patterns

Six canonical patterns for translating SportMind JSON signal outputs into visual formats. Each pattern maps to a specific output type and use case. No rendering library is prescribed — use whatever fits your stack.

Six patterns

File

platform/visual-output-patterns.md — 365 lines. Each pattern includes data input schema, full visual specification, mobile adaptation notes, and refresh cadence guidance.

Intelligence Listener

Universal monitoring layer that watches external sources across all SportMind intelligence domains and routes detected events through the three-tier classification system. The library stays zero-dependency static markdown — the listener is an optional tool in scripts/ that proposes updates.

What it monitors

Three-tier routing

Quick start

pip install requests feedparser python-dotenv

# Dry run — detect but do not dispatch
python scripts/sportmind_listener.py --dry-run

# Full cycle, all domains
python scripts/sportmind_listener.py

# One domain, webhook dispatch
python scripts/sportmind_listener.py --domain fan_token --dispatch webhook

# Custom sources
python scripts/sportmind_listener.py --custom-sources my_sources.json

Four dispatch modes

WC2026 intensive mode

The included GitHub Actions workflow (.github/workflows/intelligence-listener.yml) activates hourly monitoring of sport domain Tier 1 events June–July when WC2026_ACTIVE=true is set in repository variables. Three jobs: daily full sweep, macro every 4 hours, WC2026 hourly.

Files

scripts/sportmind_listener.py — 904 lines. 29 event types, 5 domains, 4 dispatch modes, graceful degradation without optional dependencies.
platform/intelligence-listener.md — 383 lines. Full documentation including custom source format, webhook payload schema, recommended schedules, and extension guide.
.github/workflows/intelligence-listener.yml — GitHub Actions integration with WC2026 intensive monitoring job.

Developer Toolkit

Beyond the intelligence library — ready-to-deploy templates, agent prompts, and agentic workflow patterns. These are illustrative starting points. Developers are free to create their own patterns, workflows, and integrations on top of SportMind's intelligence layer.

Copy-paste templates

Three Python files in templates/. Set your token and sport at the top, run. No architecture decisions required.

24 agent prompts

Production-ready system prompts in agent-prompts/agent-prompts.md. Copy directly into your agent's system prompt.

13 agentic workflow patterns

Full Python implementations in examples/agentic-workflows/.

78 compressed skills

Token-efficient summaries in compressed/README.md. Each compressed skill is ~500–800 tokens — approximately 70% smaller than the full skill. Use when context budget is constrained or running many simultaneous analyses.

Squad & historical intelligence

Three new core skills completing the pre-match stack. core/pre-match-squad-intelligence.md: seven-step assembly workflow with multi-sport manager language decoder (football phrases mapped to probability ranges, NBA Q/D/O system, NHL morning skate, cricket rest vs injury, MMA fight week, rugby citing). core/lineup-quality-index.md: bottom-up team strength from individual player ratings — positional weight tables for 6 sports, worked Arsenal example (LQI 0.919 vs 1.00 full-strength). core/historical-intelligence-framework.md: H2H relevance decay formula, 7 sport-specific rules, SMS-to-probability conversion, tournament bracket compounding uncertainty model.

Metric Glossary

Named composite metrics used throughout the SportMind library. Each metric is defined once and referenced across skills, MCP tools, and agent prompts.

Core signal metrics

Fan token metrics

Transfer and commercial metrics

Modifier System

SportMind signals are adjusted by a chain of multipliers. The composite modifier is the product of all individual modifiers applied to a given analysis.

Composite modifier formula

adjusted_signal = base_signal × composite_modifier

composite_modifier = macro × athlete × weather × officiating × narrative × DSM

Modifier ranges

Macro modifier values

DSM modifier values

Output Schema

Every SportMind analysis returns a consistent structured output — direction, confidence score, modifiers applied, and active flags. Defined in core/confidence-output-schema.md.

Standard signal output

{
  "direction":           "HOME",           // HOME | AWAY | DRAW | ABSTAIN
  "adjusted_score":      72.4,             // base_signal × composite_modifier
  "sms":                 79,               // SportMind Score 0–100
  "recommended_action":  "ENTER",          // ENTER | WAIT | ABSTAIN
  "composite_modifier":  1.10,

  "modifiers_applied": {
    "macro_modifier":     1.00,
    "athlete_modifier":   1.10,
    "dsm_modifier":       1.00
  },

  "flags": {
    "lineup_unconfirmed":    false,
    "macro_override_active": false,
    "citing_active":         false,
    "liquidity_warning":     false
  },

  "sportmind_version": "3.78.0"
}

Fan token signal output

{
  "token":               "PSG",
  "recommendation":      "ENTER",
  "composite_modifier":  0.98,

  "modifiers": {
    "macro":              1.00,
    "dsm":                1.00,
    "concentration":      0.90,   // from address intelligence
    "velocity":           1.03    // from address intelligence
  },

  "active_flags":        [],
  "lifecycle_phase":     3,
  "verification": {
    "chiliscan": "https://chiliscan.com/token/0xc266...",
    "fantokens": "https://www.fantokens.com/token/psg"
  }
}

ENTER / WAIT / ABSTAIN rules