RestBench — AI Restaurant Management Hackathon

Build an AI agent that runs an Italian restaurant for 30 days. Order ingredients, set prices, manage staff, run promotions — all through a REST API. The team with the highest score wins.

Your restaurant starts in the red. You have 30 simulated days to turn it around. Every decision matters: order too much and food expires, order too little and customers walk out, cut staff to save money and your reputation tanks. The best agents balance dozens of competing tradeoffs — and adapt when things go wrong.

---

Get Started (5 minutes)

1. Clone and install

git clone <this-repo>
cd restbench-starter-kit
pip install -r requirements.txt

2. Set the server URL

export RESTBENCH_URL=http://52.48.183.209:8001

Explore the API interactively: http://52.48.183.209:8001/docs (Swagger UI)

3. Run a baseline to see how the game works

python -m agents.naive_rule

This runs a simple rule-based agent that survives all 30 days but scores around -15,000. Your job: beat it.

4. Start building your agent

Option A — LLM-based agent (recommended starting point):

cp agents/llm_template.py agents/my_agent.py
export OPENAI_API_KEY=sk-...            # or ANTHROPIC_API_KEY for Claude
export AGENT_MODEL=openai/gpt-4.1-mini  # any litellm-supported model
python -m agents.my_agent

Option B — Rule-based agent:

cp agents/starter_template.py agents/my_agent.py
# Edit the strategy() function in agents/my_agent.py
python -m agents.my_agent

Option C — Any language: The API is plain HTTP + JSON. Build your agent in whatever you like.

5. Check the leaderboard

curl $RESTBENCH_URL/leaderboard

---

The Challenge

You manage a 22-table Italian restaurant. Each day, your agent receives an observation (cash, inventory, suppliers, weather, customer feedback, reputation) and responds with actions (order food, set prices, adjust staff, run promotions). After 30 days, you get a composite score.

total_score = net_profit - penalties

Penalties are assessed for low satisfaction, low reputation, walkouts, and food waste. Going bankrupt = score of -100,000.

Higher is better. The baselines score -15,000 to -19,000. A well-designed agent can score positive.

Full game specification: AGENT_CONTRACT.md Strategy thinking: STRATEGY_GUIDE.md

---

Game Loop

All interaction happens over HTTP:

POST /games                       -> create game, get first observation
loop 30 days:
    GET  /games/{id}/observe      -> read current state (optional)
    POST /games/{id}/action       -> submit one action (repeat as needed)
    POST /games/{id}/end-turn     -> advance the day, get results
GET  /games/{id}/score            -> final score

Create a game

curl -X POST $RESTBENCH_URL/games \
  -H 'Content-Type: application/json' \
  -d '{"team_name": "my_team", "scenario": "baseline", "seed": 42}'

Returns {game_id, day, status, observation}. Save the game_id.

Submit actions (one at a time, as many as you want per turn)

curl -X POST $RESTBENCH_URL/games/{game_id}/action \
  -H 'Content-Type: application/json' \
  -d '{"tool": "place_order", "args": {"supplier": "Fresh Farms NL", "ingredient": "Chicken", "quantity_kg": 8}}'

Returns {"status": "accepted"} or {"status": "rejected", "reason": "..."}.

End the turn

curl -X POST $RESTBENCH_URL/games/{game_id}/end-turn

Advances the simulation by one day. Returns the new observation plus yesterday's service results.

Get your score (after day 30 or bankruptcy)

curl $RESTBENCH_URL/games/{game_id}/score

---

Available Actions

Action	What it does
place_order	Order ingredients from a supplier. Delivery takes 1-2 days and only on that supplier's delivery days.
set_staff_level	Adjust staff between 3 and 15. Each costs 120 EUR/day. More staff = faster kitchen, fewer walkouts.
set_menu	Change active dishes (min 5). New dishes have a kitchen learning curve. Narrow menus reduce demand.
set_price	Adjust a dish's price between 0.8x and 1.2x its base price.
set_marketing_spend	Spend 0-500 EUR/day on marketing. Diminishing returns.
run_happy_hour	Boosts demand, discounts prices, small satisfaction bonus. Diminishing returns on consecutive use.
offer_daily_special	Pick one menu dish as today's special for a satisfaction bonus.
save_notes	Save up to 4,000 chars that persist between turns. Read via GET /games/{id}/notes.

Action examples (JSON)

{"tool": "place_order", "args": {"supplier": "Fresh Farms NL", "ingredient": "Chicken", "quantity_kg": 8}}
{"tool": "set_staff_level", "args": {"level": 6}}
{"tool": "set_menu", "args": {"dishes": ["Pizza Margherita", "Chicken Parmesan", "Grilled Salmon", "Mushroom Risotto", "Spaghetti Carbonara"]}}
{"tool": "set_price", "args": {"dish": "Grilled Salmon", "price": 22.0}}
{"tool": "set_marketing_spend", "args": {"amount": 200}}
{"tool": "run_happy_hour", "args": {}}
{"tool": "offer_daily_special", "args": {"dish": "Mushroom Risotto"}}
{"tool": "save_notes", "args": {"text": "Day 3: ordered salmon, low on cream"}}

---

What Your Agent Can See

The observation is returned when you create a game, call /observe, or call /end-turn.

Field	What it tells you
day, day_of_week, days_remaining	Where you are in the 30-day game
cash	Current balance (EUR)
yesterday_revenue, yesterday_total_costs	Yesterday's P&L
cost_breakdown	Staff, fixed, marketing, waste costs
inventory	Per-ingredient: total kg, batches with expiry dates, shelf life
service_summary	Yesterday's covers, revenue, walkouts, dishes sold, wait times, stockouts
supplier_catalog	All suppliers: prices, lead times, delivery days, min orders
pending_orders	Orders in transit with expected delivery day
delivery_history	Last 14 days: ordered vs delivered, on-time status
menu_book	All recipes: ingredients, base price, current price, active status
active_menu	Currently served dishes
staff_level	Current staff count
reputation_band	"Poor" / "Fair" / "Good" / "Very Good" / "Excellent"
recent_reviews	Reviews from last 14 days: stars, visit day
customer_trend	"Declining" / "Stable" / "Growing"
weather_today, weather_forecast	Today's weather + 3-day forecast (accuracy degrades)
alerts	Scenario events (supplier issues, demand changes, etc.)
notes	Your saved notes

The most important field: dishes_unavailable_at

Inside service_summary, this tells you exactly which dish ran out and when. If Grilled Salmon ran out at hour 14, you lost 8 hours of salmon sales. This is the #1 signal for inventory management.

What you DON'T see

• Exact satisfaction scores (only reputation band)
• Exact walkout counts (only band: "None" / "Few" / "Some" / "Many")
• Supplier reliability ratings
• Customer cohort sizes
• Other teams' games

---

Game Mechanics

Economics

• Starting cash: 15,000 EUR
• Daily fixed cost: 300 EUR (rent, utilities)
• Staff cost: 120 EUR/person/day (default 8 staff = 960/day)
• Total daily overhead at 8 staff: 1,260 EUR

Supply Chain

• 5-7 suppliers with different ingredients, prices, and delivery schedules
• Lead time: 1-2 days, then delivery only on supplier's specific days
• Example: Order from a Wed-only supplier on Thursday with 1-day lead = delivers next Wednesday (6 days!)
• Ingredients are perishable (3-14 day shelf life)
• Suppliers can have disruptions — deliveries may fail during outages
• Oldest batches are consumed first (FIFO)

Demand

• Varies by hour (lunch and dinner peaks), day of week (weekends busier), weather, reputation, marketing, menu variety, and pricing

Tables

• 22 tables of various sizes (2, 4, 6, 8 seats)
• Customers get the smallest table that fits; if none available, they wait briefly then leave

Reputation

• Starts at "Very Good", updated daily as a moving average
• Negative experiences have outsized impact
• Reputation spirals are real — a bad week can take many days to recover

---

Scenarios

Test your agent against different conditions. The final evaluation includes hidden scenarios your agent hasn't seen — it must adapt based on observations and alerts.

Scenario	What happens
baseline	Standard 30-day game, no events
supply_crisis	A major supplier goes into outage mid-game
tourist_season	Large demand swings: surge then drop
inflation	Ingredient prices, rent, and wages escalate over time
renovation	Reduced table capacity for first 12 days
health_scare	Viral negative reviews tank your reputation

# Play a specific scenario
curl -X POST $RESTBENCH_URL/games \
  -H 'Content-Type: application/json' \
  -d '{"team_name": "my_team", "scenario": "supply_crisis", "seed": 42}'

# List all scenarios
curl $RESTBENCH_URL/scenarios

Read the alerts. When scenario events fire, they come with alert messages in observation.alerts.

---

Tips for Winning

1. Don't run out of ingredients. A single stockout day costs revenue + reputation damage that compounds for days.
2. Watch delivery schedules. A Wed-only supplier with 1-day lead means orders placed Thursday arrive next Wednesday.
3. Check dishes_unavailable_at every turn. It's the clearest signal for what to reorder.
4. Don't double-order. Check pending_orders before placing new ones.
5. Reputation is sticky. It takes many good days to recover from one bad one. Avoid bad days rather than chasing great ones.
6. Read the alerts. "Supplier X halted operations" means find an alternative fast.
7. Use save_notes. Track orders, stockouts, and patterns. Your agent has no memory between turns otherwise.
8. Same seed + same scenario + same actions = same result. Use determinism to debug and iterate.
9. Start simple. A boring "keep everything stocked" strategy beats a clever one that occasionally runs out of food.
10. Test across scenarios. An agent that aces baseline but crashes on supply_crisis will lose on the hidden scenarios.

---

Baselines

Run these to understand the scoring range and validate your setup:

python -m agents.do_nothing        # Bankrupt by day ~16, score: -100,000
python -m agents.naive_rule         # Survives 30 days, score: ~-15,000
python -m agents.starter_template   # Rule-based starting point
python -m agents.llm_template       # LLM starting point (needs API key)
python -m agents.compare            # Run all baselines side by side

---

API Reference

Method	Endpoint	Description
POST	/games	Create a game. Body: {team_name, scenario?, seed?}
GET	/games/{id}/observe	Get current observation
POST	/games/{id}/action	Submit one action. Body: {tool, args}
POST	/games/{id}/end-turn	Advance to next day
GET	/games/{id}/score	Final score (after game ends)
GET	/games/{id}/status	Quick status: {game_id, day, cash, status}
GET	/games/{id}/notes	Read saved notes
GET	/leaderboard	Ranked scores. Filter: ?scenario=baseline
GET	/scenarios	List available scenarios
GET	/games	List games. Filter: ?team_name=my_team
DELETE	/games/{id}	Abandon a game
GET	/health	Server health check

---

Good luck. Go feed some customers.