Agent Sensei — build a bot for Ant Wars

Build a bot with your AI agent, watch it fight, and climb a live ladder.

Agent Sensei is a programming competition built for how developers actually work now: you and the coding agent of your choice — Claude Code, Cursor, Codex, … — build a bot together, then it competes bot-vs-bot on a continuous ladder. This repo is your workspace. Clone it, open it in your agent, and say "get me started."

▶ Watch live matches & read the game: agentsensei.dev/c/ant-wars · Ladder: agentsensei.dev/leaderboard

---

The game — Ant Wars

A 1v1 on a seeded, symmetric grid, under fog of war. You write the logic for one ant — and every ant in your colony runs that same program independently. There is no "command all my ants" call; coordination is emergent. Each decision sees the colony's shared sight — the union of what all your living ants can see (still fogged), so scattering scouts widens what the whole colony perceives.

• 🐜 Eat to multiply — an ant that ends its turn on food spawns a level-1 child on the next cell. Breeding never inherits power; it always starts fresh at level 1.
• ⚔️ Collide to kill — when enemy ants meet (share a cell, or swap through each other) the higher level survives, keeping its power; equal levels both die. A level-3 ant walks through level-1s untouched.
• 💪 Level up by merging — every ant has a level (power, starting at 1). MERGE fuses an ant into an adjacent ally, summing their levels into one stronger ant. A fused giant is mightier but still counts as one ant at scoring — so stacking trades numbers for muscle.
• 🧪 Talk in pheromones — drop markers (type 1–10, meaning is yours to define) that both colonies can see. Your only channel to coordinate your ants — and to bluff the enemy's.
• 🏆 Win — hold the most ants at turn ~200, or wipe the other colony out first. (A fused high-level ant counts as one, so out-numbering can beat out-muscling.)

Each turn your program reads a plain-text snapshot of what the colony can see and prints one command — CodinGame-style, no JSON:

MOVE N|E|S|W      step one cell
MOVE_TO x y       the engine pathfinds one step toward (x,y) for you
PLANT type ttl    drop a pheromone on your cell
MERGE N|E|S|W     fuse into the adjacent ally there (your levels add)
WAIT              do nothing

Full spec + the exact I/O block: .agents/game-cheatsheet.md.

---

Quickstart

The intended way — with your agent

Open this folder in your AI agent and say "get me started." It reads AGENTS.md and coaches you from zero to a ranked bot — picking a language with you, scaffolding a bot, running local matches you watch together, and explaining the game as you go. No prior game knowledge needed.

Or by hand

1. Install the sensei CLI (one binary, no dependencies):

• macOS / Linux

  curl -fsSL https://get.agentsensei.dev | sh

• Windows (PowerShell)

  irm https://get.agentsensei.dev/win | iex

• From source (any platform with Rust) — clone the platform repo and, from its engine/ directory:

  cargo build --release -p arena-cli

2. Build, play, and climb:

sensei init --lang python        # scaffold a starter bot — also: typescript | go | rust | cpp
sensei match --against baseline  # play a local match and open the replay
# ...edit bot/ until you beat the baseline...
sensei login                     # sign in (opens your browser)
sensei submit                    # put your bot on the ladder

Handy: sensei doctor checks your setup (language toolchains, editors, and whether the ladder is up) · sensei rules prints the full engine rules (--code emits the referee source) · sensei help lists every command · sensei replay latest re-opens your last match.

No account needed to build and play — everything local runs against an embedded referee that's byte-identical to the server's. You only sign in when you're ready to submit.

---

What's in this repo

Path	What it is
AGENTS.md	The coaching protocol your AI agent reads to guide you.
.agents/	Step-by-step phase guides + the game cheatsheet.
.agents/skills/	Coaching rituals (loss review, pre-submit gauntlet, boss postmortem) — cross-agent skills; mirrored in .claude/skills/ for Claude Code.
bot/	Your bot, created by sensei init. It stays on your machine — sensei submit uploads only the build artifact.
JOURNAL.md	Your engineering log, one 3-line entry per kept change — your agent drafts it, you confirm it.

Effort stats (opt-in): sensei can attach a self-reported "built with N tokens" summary to your submissions — aggregate token/model stats read from your agents' local logs, never conversation content. You'll be asked once at submit; sensei stats show prints the exact payload, sensei stats disable turns it off.

---

How it works

You iterate locally against an embedded referee that is the same code the server runs, so what you see locally is exactly what runs ranked — no drift. When you submit, your bot plays fog-of-war matches against other players on our infrastructure, an ELO-style rating settles, and every game is replayable. Climb at agentsensei.dev.