I Had a Working Agent at $3.41 a Run. Here's Where the Money Was Actually Going.

TL;DR

• Same task, same infrastructure, different harness: cost went from $3.41 to $0.03. Zero model changes.


• 
  Prompt caching on the inner browser loop (~120 lines of wrapper code): −29% cost. The 8K of repeated system prompt + tools was being billed at full price 59× per run.


• 
  Convergence rules in the prompt (three sentences): −54% cost on the desktop path, −75% wall time. The default behavior across Opus 4.8 and 4.6 is to keep searching. "After N items, STOP" is the most underrated sentence in production agent prompting.


• 
  Eliminating the LLM from the browser loop entirely (Playwright instead of browser-use): −99% cost on the browser path. The first two are optimizations within the loop. This one removes the loop.

+ markers
17:48

$2.43 (−29%)


Desktop path, base prompt
Swap browser-use for library runs a screenshot-decide-click loop inside one tool call from the outer agent. On the Yelp task, that inner loop made 59 model invocations before returning. CloudWatch's InputTokenCount for the Sonnet 4.5 inner sub-agent: 874,862 tokens in one run.

Per-call input shape:

┌──────────────────────────────────────────────────────────┐

│  Per-call input ≈ 14,800 tokens                          │

│                                                          │

│  ┌────────────────────────────────────────────┐          │

│  │ System prompt (browser-use thinking rules) │ ~5,000   │

│  ├────────────────────────────────────────────┤          │

│  │ Tool definitions (click/type/scroll/...)   │ ~3,000   │

│  ├────────────────────────────────────────────┤  IDENTICAL

│  │                                            │   every  │

│  │  ↑ 8,000 tokens repeat verbatim every turn │   turn   │

│  │                                            │          │

│  ├────────────────────────────────────────────┤          │

│  │ Conversation history (recent screenshots) │ ~5,000   │

│  ├────────────────────────────────────────────┤  varies  │

│  │ Current screenshot                         │ ~1,800   │

│  └────────────────────────────────────────────┘          │

│                                                          │

│  × 59 calls = 874K tokens billed at full price           │

└──────────────────────────────────────────────────────────┘

That ~8K of system+tools is paid at full input price 59 times. More than half the run cost is reprocessing the same instructions. Caching them is the obvious move.

Cache that, save 29%

. The general principle is well-known. What's interesting is how rarely it's applied inside agent harnesses — browser-use, series.