AGENTPITCH

01 / 13

Open Source · Python 3.11+ · Apache 2.0

AgentPitch

LLM-powered soccer simulation where every player on the field is an AI agent running a decide() callback — generated, sandboxed, and evolved by large language models.

PythonJavaScriptRust · OpenAIAnthropicGeminiCustom

github.com/gangtao/AgentPitch

Apache 2.0 — 2026

By the Numbers

02 / 13

What AgentPitch is

A complete soccer simulation engine

LLM Provider Types

Native

OpenAI

gpt-4o · o1 · o3 and latest

Native

Anthropic

claude-3.5 · claude-4 series

Native

Gemini

gemini-1.5 · gemini-2 · flash

OpenAI-Compatible

Custom

DeepSeek · OpenRouter · Ollama · any compatible endpoint

Strategy Language + Sandbox (each language ships its own runtime)

.py

Python

Sandbox

RestrictedPython

Built-in · no install · whitelist builtins

.js

JavaScript

Sandbox

QuickJS

Embedded C engine · pip install .[js]

.rs

Rust

Sandbox

Wasmtime

Compiles Rust → WASM · pip install .[wasm]

Formation Config (fully configurable)

Any N v N — set players_per_team: N in YAML config.

Tick Phases

7

Tournament Modes

3

AgentPitch · Core Metrics

02

Core Design · The decide() Interface

03 / 13

Every player is a single function

The decide( ) interface

Inputs (per tick)

game_state

Field positions · Ball location
Scores · Phase · Tick number

player_state

Position · Speed · Role
Stamina · Player ID

history

Recent actions · Last 5 decisions
Reward signal from PMEP

LLM-Generated Code

decide (

game_state,
player_state,
history

)

Python · JavaScript · Rust

Runs in sandbox · 100ms timeout

Compiled + cached between ticks

Returns (Action)

Move→ dx, dy vector

Pass→ target_player_id

Shoot→ goal direction

Tackle→ opponent_id

Hold→ stay in place

def decide(game_state, player_state, history) -> Action: # Generated by LLM · evolved by PMEP after each match

The single interface between LLM and the field engine

03

End-to-End Flow

04 / 13

How a match runs

Config → code → sandbox → play → evolve

01 · Config

YAML / API

Teams · LLM providers · Match settings

02 · CGP

Code Gen Pipeline

LLM writes decide() · Jinja2 prompt · 3 compile retries

03 · Sandbox

Compile + Cache

RestrictedPy · QuickJS · Wasmtime · 100ms timeout

04 · TickEngine

Match Simulation

Snapshot → Execute → Resolve → Physics → Log

05 · PMEP

Post-Match Evolution

LLM improves strategy from match log · top 5 events

evolved strategy fed back into next match

CGP → Sandbox → TickEngine → PMEP → loop

04

AgentPitch · Architecture

Act II · 05 / 13

Act II

Under the Hood

Four clean layers. Hard boundaries between the API surface and the simulation engine. Real sandboxes that protect the host from arbitrary LLM-generated code.

Architecture · Sandboxes · Strategy Runtime

— · —

Layer Architecture

06 / 13

Four layers · dependency: foundation → api · upper imports lower

04
TOP

API Layer

FastAPI (HTTP server) · React (browser UI) · SSE (server-sent events, live stream) · Pydantic (independent API models)

03

Orchestration Layer

TE (TickEngine) · ARE (ActionResolutionEngine) · CLI (command-line runners: season / cup / league)

02

Core Layer

GSM (GameStateManager) · PMS (PlayerMovementSystem) · BPS (BallPhysicsSystem) · MLS (MatchLogSystem)

01
BASE

Foundation Layer

PAL (Provider Abstraction Layer) · SF (SandboxFactory) · CGP (Code Generation Pipeline) · PMEP (Post-Match Evolution Pipeline) · ARE (ActionResolutionEngine) · GSS (Game State Schema)

Upper layers import lower — not the reverse

Foundation · Core · Orchestration · API

06

TickEngine · Per-Tick Resolution Pipeline

07 / 13

What happens inside every single game tick

7 tick phases

Phase 01

1

Snapshot Collection

Capture full game state from GSM — positions, scores, ball, phase, tick index.

GSM

Phase 02

2

Action Generation

Invoke each player's decide() in sandbox. Failures routed to FallbackHandler.

Sandbox

Phase 03

3

Validation & Normalization

Cooldown checks, Move speed clamping, Pass/Shoot power capping, Tackle target validation.

Validate

Phase 04

4

Player Movement

Compute-all-then-commit. PMS resolves Move actions, player separation, dribble contests.

PMS

Phase 05

5

Ball Actions

Pass & Shoot resolution. Set ball velocity, landing zone, skill-based deviation, transfer possession.

Ball

Phase 06

6

Tackle Resolution

Range check, possession verification, strength-based success probability per Tackle action.

Contest

Phase 07

7

Ball Physics & Goal

BPS advances ball. Goal-line crossing, scoring, and goalkeeper save attempts resolved.

BPS · Goal

Phases execute sequentially · results merged into action_records dict · logged to MatchLogSystem ~100ms timeout per decide()

07

Strategy Runtime · Language & Provider Support

08 / 13

Write code · run in a sandbox · evolved after every match

Three strategy languages

Language	Sandbox Backend	Security Model
Python	RestrictedPython	Whitelist builtins · no imports · exec isolated
JavaScript	QuickJS (embedded C engine)	Isolated JS runtime · no Node APIs
Rust → WASM	Wasmtime	WASM sandbox · memory-isolated · AOT compiled

LLM Providers (via PAL — Provider Abstraction Layer)

openai/

OpenAI

gpt-4o · o1 · o3

anthropic/

Anthropic

claude-3.5 · claude-4

google/

Gemini

gemini-1.5 · gemini-2

deepseek/

DeepSeek

deepseek-v3 · v4

openrouter/

OpenRouter

Any hosted model

local/

Ollama

Fully offline · no API key

Safety Guarantee

LLM-generated code never touches the host filesystem, network, or OS — it lives entirely inside the sandbox with a 100ms execution budget per tick.

All code runs in a language-appropriate sandbox — no host access

08

Browser UI · Live Match Viewer

09 / 13

Real-time simulation · browser on port 8765

Live match viewer

Field · Live ViewSVG canvas · real-time

5v5 field in SVG. Player positions and ball updated every tick via SSE stream.

Event FeedColor-coded by type

Goal · Shot · Pass · Tackle events in scrollable log. Click any event to scrub to that tick.

Stats PanelPost-match breakdown

Possession · shots · passes · tackles per player. Tabular monospace numerics.

FastAPI + React · SSE for live events · port 8765

08

AgentPitch · Tournament Modes

Act III · 09 / 13

Act III

Tournaments

Three structured formats for AI competition. Every match feeds back into strategy evolution — the longer the tournament, the smarter the agents become.

Arena · Cup · League

— · —

Arena Mode · Head-to-Head Evolution

11 / 13

Arena · two LLMs · configurable match series · default 3 matches · one evolving rivalry