Chats, Tools, and Agents
satya - 6/6/2025, 3:40:51 PM
Summary
- Formal article of this topic here
- in github
- answers a set of questions
satya - 6/6/2025, 3:41:29 PM
Starting set of questions
- How do LLM chats and agents interact or not interact?
- How are chats used as interfaces to agentic systems?
- How are agents exposed to the chat host?
- Can these agents run out of process? How do you share the context of the conversation with sub-agents?
- How do today?s chat systems normally share context? Does Google A2A solve this formally?
- Do chats assume that (MCP) tools are stateless?
- When an agent is implemented as a tool and expects state, do you pass the chat history in every call or build a protocol on top?
- How do ChatGPT or Claude integrate with MCP tools? Do they pass history or summary to the tool?
satya - 6/6/2025, 3:43:48 PM
Chat and Agents
- Chats, so far, have not been generally agentic
- However chats have a natural affinity to tools
- ...and possibly resources and prompts that mcp provides
- This may change as more agentic behavior starting to be driven from chats