What a harness is — the four layers under every coding agent — step 9 of 9

Final drill. Build a harness with all four layers measurable. Write harness(user_input, tools, fake_model, config, max_iters) that:

• Layer 1: hold system from config in a separate system_prompt variable (matches Anthropic's SDK shape: system= is a top-level kwarg, NOT a role inside messages). Build messages with just the user turn.
• Loop up to max_iters. Layer 2: call fake_model(messages) via with_retries (provided — retries on TransientError). Layer 3: filter content into text + tool_use blocks; track tokens_used from response.get("usage", 0). On end_turn: return a result dict including layer counters. Layer 4: dispatch each tool_use block; track tool_calls_total.
• Returns {"text": str, "iters": int, "tool_calls": int, "tokens": int, "system_used": bool}.
• system_used is True iff the config's system key was held in system_prompt (i.e., would be sent via the SDK's top-level system= kwarg).

Two cases run. Expected output:

text='Found ramen.' iters=2 tool_calls=1 tokens=130 system_used=True
text='capped' iters=2 tool_calls=2 tokens=180 system_used=False