Stanford Online

How AI Agents Work: Connectors, MCP & Agent Safety | Stanford CS193T Lecture 5 with Cynthia Bailey: summary

YouTube summary5 sectionsWatch on YouTube ↗

This is an AI-generated summary of the YouTube video "How AI Agents Work: Connectors, MCP & Agent Safety | Stanford CS193T Lecture 5 with Cynthia Bailey" (Stanford Online), made with Samuraize and published by Samuraize. It condenses the YouTube video into 5 titled sections you can read in a couple of minutes, each linking to the moment in the video it covers.

1
Filed under💻 Technology0 comments🍱 Add to trayReport
Study this

How AI Agents Work: Connectors, MCP & Agent Safety | Stanford CS193T Lecture 5 with Cynthia Bailey

Stanford Online

切

Connectors and action flow 0:05

Connectors are access permissions that let an AI tool read data from other apps and act in them. With email, you can either copy messages into chat and paste revisions back yourself, or let the model read your inbox, draft a reply, and even send it. The key shift is from manual content help to action-taking, while the LLM still produces text one token at a time.

切

MCP and agent setup 17:00

That action path is called the model context protocol, or MCP. A connector manager reads structured output, usually JSON, and turns it into tool calls, then feeds the results back into context. To design an agent, you specify its goal, context, memories, tools, and permissions. A dinner-scheduling agent may need calendar access, but not every agent in the same workflow should have the same powers.

切

Agent loops and stopping 21:02

Agents need a clear task, the context they should use, the tools they can call, the permissions they have, the memory they can keep, the actions they should take, and the stopping conditions. For the team dinner example, that means calendars, messaging, past scheduling preferences, and a step-by-step path: check availability, draft a message, send it, then gather replies. If the dinner cannot be set after a few tries, the agent should stop and report the problem.

切

Analyze act observe 24:00

Agent work is built around an analyze-act-observe loop. The model looks at the context, decides what to do, performs an action, and then checks whether it worked and had the right effect. That matters because model output is probabilistic. A JSON tool call can come back slightly wrong, such as saying check weather instead of get weather, and then the connector will fail. Human in the loop means deciding where people should see, approve, or gatekeep progress.

切

Agent safety 41:31

The discussion shifts from connectors to safety and asks you to think like a pessimistic chief information security officer. Three agent setups are tested for the worst case. A dorm leak agent can read photos on your phone and send an Outlook report with maintenance details. A job hunt agent can read and write LinkedIn messages and use web search to reply to recruiters in an informed way. A disk cleanup agent can read files on your laptop and permanently delete anything untouched for six months until storage drops to a target level. Each setup exposes the same question: what can go wrong when the agent has broad access, repeats actions, or is given delete power?

AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

Summarize your own YouTube video

Paste a YouTube link, article, PDF, ebook or slide deck and get a summary like this in seconds. Free to try, no sign-up needed.

⚔️ Try the YouTube summarizer

Discussion

Sign in to join the discussion. Sign in

More from the Bento Box

Browse the Bento Box →

With your OK, Microsoft Clarity and Google Analytics set cookies (data goes to the US) so we can see what breaks. Cookie details

Samuraize.ai - How AI Agents Work: Connectors, MCP & Agent Safety | Stanford CS193T Lecture 5 with Cynthia Bailey