Click any words you think contain an error. Click again to unmark.
Mark at least one word to submit.
The bug is in the line that says the tool returns the entire 80KB HTML page and the agent appends it verbatim. That single design choice causes two compounding problems.
First, raw HTML is almost all chrome and almost no content. A typical web page is mostly tags, inline scripts, navigation menus, ads, and footer boilerplate. The 80KB page might have 2KB of actual answer-relevant text, so 78KB out of 80KB is token cost with no information value.
Second, once that blob is in the assistant turn, it rides along on every subsequent call in the same conversation. Pay once at iteration 1, pay again at iteration 2, pay again at iteration 3, and so on. The structural fix is to extract the relevant fields (title, snippet, URL, a body excerpt) at the tool wrapper layer, cap each field at a known size, and return a small structured object. The raw HTML never reaches the prompt.
Two failures at once. The error span is 'returns the entire 80KB HTML page and the agent appends it verbatim', that is the bug. The fix is to move the extraction and capping logic into the tool wrapper, before the result enters the message stream.
A correct search tool returns a small structured object: title, URL, snippet, perhaps a short body excerpt of a few hundred characters, and a reference id pointing to the full raw payload stashed out of band. The model gets enough signal to decide whether the result is relevant, and the agent has a path to ask for more if it needs it. The 80KB raw page never enters the prompt.
The compounding factor is the agent loop. A single 80KB blob hurts once; carried forward across ten iterations it accumulates into roughly 200K tokens of input billed per call and degrades the lost-in-the-middle behaviour for the user's actual question. Tools that return raw payloads are the single biggest source of context bloat in production agents in 2026.
Quick reference
TL;DR
Raw 80KB HTML pasted verbatim into the assistant turn is the bug, the tool wrapper should parse, extract, and cap before anything reaches the prompt.
Imagine asking a friend to summarise a long Wikipedia article and they hand you the whole printout including the sidebar, the ads, the page footer, the citation list, and the edit history. You wanted the answer; you got a stack of paper. Worse, every time you ask a follow-up they hand you the same stack again because they keep it in their lap. The tool wrapper is the friend's job, read once, take out the useful paragraph, throw the rest away, and only hand you the small note.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Tools that already return bounded structured payloads by design (calculators, single-row lookups) The extraction layer adds complexity for no win when the natural output is already small.
Auditing or compliance flows that require the raw payload as evidence In those flows the raw blob is the artifact, not the context, it lives in an audit store, not the prompt.
One-shot calls where the agent will not iterate further The accumulation problem disappears for non-iterative calls; the precision loss from extraction may not be worth the complexity.
Quick scan · 2–4 min
Everything important, quickly.
7 min: identifying the bug span, why HTML is mostly chrome, the parse extract cap reference pattern, and how the fix preserves prompt caching and lost-in-the-middle behaviour.
Code example
python
# Before: raw HTML to context (the bug in the question)
result = http_get(url).text # 80KB of HTML
messages.append({"role": "tool", "content": result})
# After: parse, extract, cap, reference
from readability import Document
import hashlib
MAX_EXCERPT_CHARS = 800 # roughly 200 tokens
def search_wrapper(url: str, store) -> dict:
raw = http_get(url).text
doc = Document(raw)
body = doc.summary(html_partial=False)
excerpt = body[:MAX_EXCERPT_CHARS]
ref_id = hashlib.sha1(url.encode()).hexdigest()[:10]
store.put(ref_id, raw)
return {
"title": doc.title(),
"url": url,
"excerpt": excerpt,
"truncated": len(body) > MAX_EXCERPT_CHARS,
"ref_id": ref_id,
}
Real-world examples
Real products, models, and research that use this idea.
1Anthropic's web_search tool returns title, URL, and a short snippet, never the full HTML, exactly to avoid this failure.
2Perplexity's API returns structured search results with bounded fields per source.
3Mozilla Readability and trafilatura are the standard parsers production teams use to extract main content before injection.
4Claude Code's WebFetch tool runs an LLM-summarisation pass on the page before the result reaches the agent loop.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you handle a tool that returns JSON with deeply nested objects rather than HTML?
Sign in or go Pro to see the approach for this follow-up.
QWhat if the agent genuinely needs to read the full page for one specific call?
Sign in or go Pro to see the approach for this follow-up.
QHow does this interact with prompt caching on Anthropic or OpenAI?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating the model as the place to fix verbose tool outputs, the model has no control over its own inputs; the tool wrapper does.