An agent can receive a CAPTCHA, an empty page, fake content, or stale training data and still answer as if its web request succeeded.
2
In Rafael Levi's comparison, identical tasks produced zero successes without Bright Data MCP, while the MCP run reached several properties and social and shopping pages.
3
Agents should use only the MCP tools they need, since loading all available tools floods the context with irrelevant information.
Summary
Rafael Levi describes a failure mode where an agent cannot access a website but hides that failure behind a confident answer. Anti-bot systems can return CAPTCHAs, empty pages, or misleading content, while the model may fall back to training data and present it as current. Levi demonstrates the problem with identical prompts for property listings, LinkedIn, Instagram, Amazon, and TikTok. Without web tools, all five tasks fail. With Bright Data's Web MCP, the run succeeds on several of them. He explains that the MCP provides search, markdown scraping, batch search, pre-built website APIs, and a remote browser with CAPTCHA handling. The talk also sets limits: Bright Data works with publicly available data and does not collect data behind login. Levi recommends loading only the tools an agent needs. He also describes using skills to have an agent build a scraper and parser, rather than asking an LLM to parse every page itself.
An agent can hide a failed web request behind a confident answer
Rafael Levi says LLMs are programmed to please users, so they often keep trying when they cannot access requested data. A blocked request may produce a CAPTCHA or an empty page without an explicit error. The agent then makes up an answer, sometimes with numbers or citations that do not work. Levi describes this as an invisible failure because there is no warning for the user to inspect. He would prefer the model to say, "No, I can't," but says it usually does not.
Anti-bot systems can make an agent's data look current when it is not
Levi says websites have been fighting robots and automation for years, and that AI systems now face AI-oriented blocking. When a request fails, an agent may fall back to training data and describe it as the current situation. He contrasts training data from 2024 with the talk's 2026 setting. He also describes Cloudflare's AI Labyrinth as a system that can feed bots fake data instead of simply blocking them. That creates a harder problem because the result can look plausible while being wrong.
The demo compares identical tasks with and without live web tools
Levi runs the same prompts in two scripts, one without MCP and one with Bright Data MCP. The tasks target properties on Rightmove, a LinkedIn company, an Instagram account, an Amazon product, and a TikTok page. Without MCP, the model has no live web access or browsing tools available by default, and the displayed result is zero successes and five failures. With the MCP, the run succeeds on Rightmove, LinkedIn, Instagram, and the Amazon product. Levi then asks the model to compare both sets of results.
Bright Data MCP combines search, scraping, and browser access
Levi describes an MCP with search for Google, Bing, and DuckDuckGo, plus a markdown scraper that returns page content without HTML tags. Batch search can send many keywords together. The system also includes pre-built APIs for some websites and a remote browser that an LLM can open and navigate. Levi says the browser has a unique fingerprint and can handle CAPTCHAs, allowing multiple sessions to run in parallel. The demo itself used direct URL access rather than search, although he says the MCP can also perform Google searches.
Levi says Bright Data does not collect data behind login. He frames this as a legal boundary because users accept a site's terms and conditions when they create an account, and those terms may restrict scraping or robot access. He demonstrates public LinkedIn data in an incognito window, while explaining that access can vary by IP quality and request volume. A home connection may show several profiles before requiring a login. For data that does not need to be live, he also describes datasets that agents can filter, such as LinkedIn people by role and location.
Making the agent look human is Bright Data's answer to blocking
When asked how the system handles Cloudflare's misleading pages, Levi says the approach is to make the agent appear as human as possible. The browser uses recorded mouse movement and typing that mimics human behavior. He says this can avoid triggering a block without needing to understand exactly how the site detects bots. He is less certain about misleading content. Hotels may show different prices to a phone, a computer, or a proxy, and Levi says the practical approach is to make the agent look human and hope for the best.
Tool selection matters because an MCP can expose dozens of tools
In the questions, Levi is asked whether an agent must load all the tools exposed by the MCP. He says no. If an agent only needs markdown scraping and search, it should load those two tools. Loading the full set would flood the context with irrelevant information. He also describes a skills page that teaches an agent how to build a scraper or data pipeline. For large collections, his recommended pattern is to have the LLM build a parser once and let a script run it, rather than asking the LLM to parse every individual HTML page.
"The invisible failure group, right? There's no error, no warning, just the wrong answer, right?"01:55
Who should watch
You are building an agent that gives product links, profiles, prices, or other supposedly current web data, and you need to detect silent access failures.
Your agent can browse in theory but often returns stale facts, broken citations, or plausible pages after a CAPTCHA or blocked request.
You are adding an MCP with many tools and need a practical rule for keeping irrelevant tool definitions out of the agent's context.
