Agents plan actions, use tools, and reason over results until they reach an objective.
2
Amazon Nova Act lets Python applications control browsers with natural-language instructions, including parallel web tasks.
3
MCP connects Nova Act and other tools to clients such as Amazon Q, Cursor, and Claude Desktop, while Strands provides a code-first way to compose agents.
Summary
Du'An Lightfoot and Banjo Bami introduce a practical agent architecture built around planning, tool use, and reasoning. The workshop then moves through three modules. Participants call Amazon Nova Act from Python to search Amazon and extract product information, including parallel searches across monitors. They wrap Nova Act in an MCP server so an LLM client can translate natural-language requests into browser actions. Finally, they use the open-source Strands framework to combine prompts, models, tools, MCP servers, and multiple agents. The demos show where the approach works and where it fails. Nova Act can interact with sites that lack APIs, but it is a research preview, may get stuck, and cannot bypass CAPTCHA challenges. The speakers also explain authentication, human pauses, browser profiles, model selection, and the difference between Nova Act's browser model and the Bedrock model used by an MCP client.
Lightfoot defines an agent around three steps. It receives a prompt and objective, determines the actions required, and creates a plan. It then uses tools to carry out those actions. The agent evaluates the results and decides whether to update the plan or take more actions until the objective is complete. The surrounding architecture can include an LLM, a knowledge base, guardrails, tools, memory, human input, and calls to other agents or models such as Amazon Nova Act through MCP.
The speakers describe evaluation as an ongoing part of building agentic systems. Teams need to test whether they chose the right LLM and whether prompts are consistent, accurate, and optimized for the expected performance. They also need a way to judge whether the system is solving the intended problem. Logs provide the evidence for that review. A subject matter expert can then inspect the results and decide how to improve the system through repeated iteration.
Simple tasks should stay with ordinary program logic
Agentic systems fit tasks where the developer does not know in advance which tools or how many tools will be needed, and where the model's reasoning can help choose the next step. A one-step operation does not need that machinery. Lightfoot recommends a traditional 'if this then that' solution for those cases. He presents AWS agent options ranging from specialized tools such as Amazon Q, to managed Amazon Bedrock Agents, to a do-it-yourself approach using the open-source Strands framework.
Nova Act turns natural-language instructions into browser actions
The first Nova Act exercise opens Amazon.com, searches for a coffee maker, selects the first result, and returns the product title. The Python code describes the task without locating individual HTML tags, divs, or H1 elements. Nova Act sees the page and acts on elements such as the search box and buttons. The run produces screenshots, an action log, and a video recording of the browser session. The demo returns a Black+Decker 12-cup coffee maker title.
Parallel browser sessions can collect comparable results
A second example searches for several monitors at the same time. The code specifies the fields to extract, including price, rating, and size, then starts separate threads for the monitor tasks. Nova Act runs the sessions in headless mode and returns the collected information without making the audience wait through every browser action. Lightfoot suggests the same pattern for recurring research tasks on sites that do not provide a native API, such as gathering daily news.
Nova Act has clear limits around CAPTCHA and ambiguous workflows
Nova Act is a research preview and the demonstrations expose its boundaries. It cannot bypass CAPTCHA challenges, and the speaker shows it getting stuck when Google presents one. It can loop while scrolling and may need more explicit instructions on unfamiliar or intricate sites. Human involvement is possible through Python code that pauses for credentials, passwords, or other input, and a local browser can remain visible. The model can use an existing browser session, but two-factor authentication and CAPTCHA remain difficult cases.
Lightfoot's Nova Act MCP server exposes tools such as browser sessions, browser actions, parallel tasks, screenshots, result listing, and closing the browser. A client such as Amazon Q, Cursor, or Claude Desktop can receive a natural-language request and generate the Nova Act code needed to execute it. The MCP client still needs an LLM, such as Claude 3.5 Sonnet through Bedrock, to interpret the request and select tools. The server provides the interface that translates those tool calls into browser actions.
Strands composes models, tools, prompts, and other agents
Strands is presented as an open-source, model-first agent framework that can be installed with pip. A basic agent combines a model, a system prompt, and tools. It has native MCP support and can use Amazon Bedrock, LiteLLM, or Ollama, so it is not limited to AWS models or services. The workshop example uses AWS documentation and AWS diagrams MCP servers to find Lambda documentation and create an architecture diagram. Separate agents can have their own prompts, models, and tools, then be called by an orchestrator.
"We don't always need to provide some type of agentic system for something that can be done with a traditional solution."04:20
Who should watch
You are building an agent that needs to browse websites or work with interfaces that do not expose an API.
You want to connect browser automation and external tools to clients such as Amazon Q or Cursor through MCP.
You are comparing managed Bedrock Agents with a code-first framework and want to see how Strands handles MCP and multi-agent workflows.
