SentienceAPI Snapshot Experiment with browser-use Agent

[rcholic]

Snapshot Elements Are Formatted For LLM Agent Reasoning

Element ID	Role	Name	Importance	doc_y	ordinal	dominant group	href
783	textbox		1014	6	-	0
16	link	Hacker News	307	0	7	1	ycombinator
454	link	Show HN: InfiniteGPU, An op...	192	3	230	1	github
550	link	Show HN: ElixirBrowser – An...	189	4	282	1	github
289	link	Show HN: Miditui – A termin...	188	2	141	1	github
313	link	Show HN: EuConform – Offlin...	184	2	154	1	github
622	link	Show HN: Roleplay-first cha...	184	5	320	1	abliteration
526	link	Show HN: I built a Postgres...	183	4	269	1	postgresgui
646	link	Show HN: I vibecoded an ARM...	183	5	333	1	github
742	link	Show HN: SMTP Tunnel – A SO...	181	6	385	1	github
430	link	Show HN: Repogen – a static...	178	3	217	1	github
406	link	Show HN: macOS menu bar app...	177	3	205	1	github
217	link	Show HN: Similarity = cosin...	177	2	104	1	github
478	link	Show HN : A game to documen...	175	4	243	1	jeevan
241	link	Show HN: I made a memory ga...	175	2	116	1	specr
718	link	Show HN: A geofence-based s...	175	5	372	1	localvideoapp
169	link	Show HN: Why Nvidia's PhysX...	175	1	78	1	github
598	link	Show HN: I visualized the e...	174	4	308	1	bikemap
502	link	Show HN: I built a tool to ...	174	4	256	1	tryflux
49* (top)	link	Show HN: 15 Years of StarCr...	173	0	15	1	migdal
121	link	Show HN: Agent-of-empires: ...	173	1	53	1	github
670	link	Show HN: I built a "Do not ...	168	5	346	1	apoorv
574	link	Show HN: DeepDream for Vide...	168	4	295	1	github
73	link	Show HN: Yuanzai World – LL...	165	1	28	1	yuanzai

INFO     [state_injector] ✅ Sentience snapshot: 50 elements, URL: https://news.ycombinator.com/show
INFO     [Agent] 🧠 Sentience: Injected 50 semantic elements (1557 chars) into LLM context
INFO     [service] ✅ Vision is DISABLED: use_vision=False, screenshots=0, sample_images=0, read_state_images=0
INFO     [service] ✅ Sentience detected - reducing DOM size to 5000 chars
INFO     [prompts] 📊 DOM state: 10412 chars (~2603 tokens) truncated to 5000 chars (~1250 tokens)
INFO     [prompts] 📊 DOM state: 10412 chars (~2603 tokens) truncated to 5000 chars (~1250 tokens)
INFO     [prompts] 📊 Token breakdown (chars): agent_history=680 (~170 tokens), browser_state=5414 (~1353 tokens), agent_state=228 (~57 tokens), read_state=0 (~0 tokens), total=6418 (~1604 tokens)
INFO     [prompts] ✅ AgentMessagePrompt: Vision DISABLED - use_vision=False, screenshots=0, sample_images=0, read_state_images=0
INFO     [Agent]   🧠 Memory: I have successfully navigated to the "Show HN" page. The goal is to find the number 1 post. Looking at the browser state, the post ranked 1 is clearly visible. The title is "Show HN: 15 Years of StarCraft II Balance Changes Visualized Interactively". I will extract this title and use the 'done' action to complete the task.
INFO     [Agent]   ▶️   done: text: The number 1 post on Show HN is: Show HN: 15 Years of StarCraft II Balance Changes Visualized Interactively, success: True

📄 Final Result:
The number 1 post on Show HN is: Show HN: 15 Years of StarCraft II Balance Changes Visualized Interactively

INFO     [Agent] ✅ Task completed successfully
INFO     [BrowserSession] 📢 on_BrowserStopEvent - Calling reset() (force=True, keep_alive=None)
INFO     [BrowserSession] [SessionManager] Cleared all owned data (targets, sessions, mappings)
INFO     [BrowserSession] ✅ Browser session reset complete
INFO     [BrowserSession] ✅ Browser session reset complete

Token Usage Comparison:

Before (i.e. Use Vision Model & Full DOM)

TLDR;

DOM limit: 40,000 chars;
DOM state: 10,195 chars (~2548 tokens) - not truncated;
Total per step: ~12,667 chars (~3,166 tokens)

Overall Token Usage Summary

Metric	Value
Total Prompt Tokens	35,740
Total Prompt Cost	$0.006997
Total Prompt Cached Tokens	839
Total Prompt Cached Cost	$0.0000168
Total Completion Tokens	1,311
Total Completion Cost	$0.002622
Total Tokens	37,051
Total Cost	$0.009636
Entry Count	7

Per-Model Breakdown (bu-1-0)

Metric	Value
Model	bu-1-0
Prompt Tokens	35,740
Completion Tokens	1,311
Total Tokens	37,051
Cost	$0.009619
Invocations	7
Average Tokens per Invocation	5,293.0

Cost Breakdown

Category	Tokens	Cost
Prompt (uncached)	34,901	$0.006997
Prompt (cached)	839	$0.0000168
Completion	1,311	$0.002622
Total	37,051	$0.009636

After (i.e. Use SentienceAPI SDK with ranked DOM)

TLDR;

DOM limit: 5,000 chars;
DOM state: 5,000 chars (~1,250 tokens) - saved 1,353 tokens;
Total per step: ~6,418 (~1,604 tokens) - 50% reduction

Overall Token Usage Summary

Metric	Value
Total Prompt Tokens	13,251
Total Prompt Cost	$0.0025026
Total Prompt Cached Tokens	820
Total Prompt Cached Cost	$0.0000164
Total Completion Tokens	892
Total Completion Cost	$0.001784
Total Tokens	14,143
Total Cost	$0.004303
Entry Count	4

Per-Model Breakdown (bu-1-0)

Metric	Value
Model	bu-1-0
Prompt Tokens	13,251
Completion Tokens	892
Total Tokens	14,143
Cost	$0.004287
Invocations	4
Average Tokens per Invocation	3,535.75

Cost Breakdown

Category	Tokens	Cost
Prompt (uncached)	12,431	$0.0025026
Prompt (cached)	820	$0.0000164
Completion	892	$0.001784
Total	14,143	$0.004303

Overlay During Agent Runtime

Improvements with SentienceAPI SDK

1. Vision disabled — no screenshots/images sent to LLM
2. DOM truncation working — 10,412 chars → 5,000 chars (saves ~1,353 tokens per step)
3. Task completed successfully — agent found the correct top post
4. Token usage is roughly halved per step while maintaining accuracy. The agent is using Sentience semantic geometry instead of full DOM + screenshots, which is more efficient & cost-effective.