Exploring Agent Tech Stack in 2025: Simplifying Agents with Model Context Protocol (MCP)

Join Sarmad, CEO of Last Mile AI, as he explores the exciting possibilities of agent technology and shares his insights on building effective agents with Model Context Protocol (MCP) in 2025.

  • Speaker is Sarmad, CEO of Last Mile AI and previously worked on Language Server Protocol (LSP) at Microsoft.
  • Believes 2025 will be the year of agents, with Model Context Protocol (MCP) making agent design simpler and more robust.
  • MCP servers can be more than just one-to-one mappings of existing REST API services.
  • Agents can be represented as MCP servers, allowing for asynchronous workflows using infrastructure like Airflow or Temporal.
  • LSP revolutionized the developer tools ecosystem by standardizing a single interface for language services in IDEs.
  • Model Context Protocol (MCP) is a similar effort to standardize a single interface for connecting LLMs to tools, data, and resources.
  • MCP has been adopted by major companies like Google, OpenAI, and Microsoft, making it the de facto standard for connecting LLMs to the world around them.
  • Building effective agents in 2025 is made possible by better models, MCP, and simpler agent application architectures.
  • Better models include reasoning models and large language models (LLMs) that are reliable for various use cases.
  • Test time compute has enabled more complex patterns like chain of thought reasoning or react to shift from the framework layer to the inference layer, reducing app developer burden.
  • Agents are now simply orchestrators of better models and MCP, with no need for monolithic AI frameworks.
  • Anthropic's blog post "Building Effective Agents" highlights successful agent patterns, such as the augmented LLM and optimizer/evaluator LLM model.
  • Sarmad built an MCP-native agent library called MCP Agent during his Christmas break, implementing various agent patterns from Anthropic's blog post.
  • The benefits of exposing agents as MCP servers include composable, platform-agnostic, and scalable agents.
  • Agents can be modeled as asynchronous workflows, allowing for pausing, resuming, retrying, triggering, scheduling, and human feedback.
  • MCP Agent uses Temporal as the durable execution backend for agent orchestration and compute.
  • A demo is provided to showcase an agent loading a student's short story from a markdown file, analyzing it, and generating a graded report using various MCP servers and agents.
  • Asynchronous agents allow users to kick off agent tasks from anywhere and check their status at a later time.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!