F1 Race Strategy Agent — AI skill for Claude Code

F1 Race Strategy Engineer Agent — CSE 598 Capstone Baseline

An **agentic** baseline: Claude acts as a Formula 1 race strategist. Given a driver and their starting tyre, the agent reasons over **real 2021 Spanish GP telemetry** to recommend a pit/tyre strategy. It runs a genuine *plan → call tools → observe → refine* loop — it inspects tyre degradation, proposes candidate one- and two-stop strategies, simulates each with a deterministic race-time model, compares them, and submits a final recommendation.

The recommendation is **objectively checkable**: it is compared against the optimum of the same deterministic simulator (see `data/ground_truth.json`).

What the baseline does (step by step)

1. Model — Anthropic claude-opus-5 (configurable) via the official anthropic SDK, driven by a hand-written tool-use loop in `src/agent.py` (the loop is explicit on purpose, so the agentic cycle is visible).
2. Tools the agent can call (`src/tools.py`):
   • get_race_info — race length, compounds, pit-loss, min stint length.
   • get_driver_stints — a driver's actual stints + starting compound (from data).
   • estimate_tyre_degradation — fits a linear pace-vs-tyre-age model per compound from that race's green-flag laps.
   • simulate_strategy — deterministic estimate of total race time for any proposed stint plan (validates: laps sum to race length, ≥2 compounds, min stint length).
3. Data — real timing data from FastF1, pre-downloaded into cache/ and read in offline mode → no network needed for race data.

Setup

git clone https://github.com/DoomDust7/f1-race-strategy-agent.git
cd f1-race-strategy-agent
python -m venv .venv
source .venv/bin/activate           # Windows: .venv\Scripts\activate
pip install -r requirements.txt

API key (required)

The agent calls the Anthropic API, so you need an API key with credits.

cp .env.example .