View this page on GitHub

Expanding on notes captured on the Glossary page

0️⃣ Links to Check Out

1️⃣ A2A (Agent-to-Agent) — The Intent Layer

  • An MCP server exposes:
    • tools (functions the client/LLM can call)
    • resources (documents, URLs, databases)
    • prompts (template suggestions)
    • context extensions (metadata, workspace awareness)

2️⃣ Prediction Guard’s Agent “Chunks”

From screenshots captured in a webinar, here’s how Prediction Guard chunks up the components of an Agent (which may/do integrate to MCP servers)

System Prompt

Example: Hello! I'm your Go-to-Market specialist from FutureShield. How can I assist you today? Whether you need insights on our sales pipeline, upcoming contract renewals, or anything else to help drive growth, just let me know!

MCP Server & Tools

(Assume this would be specified… get the slides from the Prediction Guard guys)

Settings

  • Temperature: a real number between 0 and 1 that controls how much randomness there is in the model’s responses (0 is most predictable/least variable [creative])
  • Top P: Top-P (also called nucleus sampling) limits the model’s next-token choices to the smallest set of tokens whose cumulative probability adds up to P (e.g. 0.80)
  • Top K: Top-K limits the model’s next-token choice to the K most probable tokens, then samples from only that fixed set
  • Max Output Tokens: Maximum number of tokens the model can generate in a single response

Resources & Knowledge

  • Additional Documents: assume for RAG purposes e.g. uploaded documents

Security & Safety

(What tools can be used to execute on these guardrails?)

  • Prompt Injection Protection
  • PII Detection & Replacement

Security considerations:

  • High level MCP Governance:
    • Leverage monitoring
    • Employ:
      • MCP server “allow” lists
      • MCP server and tool “verified” lists
      • MCP server “blacklists”
  • Current standard is to use OAuth 2.1 authentication to authenticate your MCP client to MCP servers
  • Apply prompt injection detection to MCP server results to safeguard against adversarial or malicious inputs

Other Notes & Ideas

  • IDEA: build a RAG system and wrap in an MCP server so it can be integrated into Agents
View this page on GitHub