Recommended tool for an agent for a specific use case
Here is the use case:
A Commercial Agent (phase 1)
Step 1 - Trigger - Slack form submitted and sent to Claude
Step 2 - Agent in Claude - will pull data from Snowflake and Salesforce to come up with the output
Step 3 - The agent will push the results to Workato
Step 4 - Workato will push the result to the opportunity in Salesforce
Architecture query: Is it possible to run a automated workflow and agent in Claude itself? Or would we need to utilise n8n for the automated orchestration layer
ie Is this possible?
Can the agent sit in claude (with system prompt and instructions) and is triggered each time a new Slack submission is submitted and pushed to the agent (via the Slack to Claude MCP)
Tools attached to the agent ( Snowflake and SF MCP connectors to Claude) are called and used by the instructions
Output from the agent is a JSON sent to Workato from Claude (automatically)
Workato does its magic (requirement of Systems team if writing to Production environment in Salesforce) and writes to Salesforce
tbh u want both. claude’s API is reactive — u call it, it responds. it doesnt subscribe to slack or fire on a schedule on its own, so something has to be listening to slack and calling claude. n8n’s a clean fit bc the Slack Trigger node is built in and the Anthropic node lives in the same workflow, so retry / error / logs come for free.
the “agent in claude itself” piece is the one part u cant do yet. the anthropic API has no equivalent of “subscribe to my slack workspace,” and the slack MCP servers are for conversational chat in claude.ai, not workflow triggers. so the trigger has to live somewhere outside claude.
if u dont want n8n in the loop, alternatives that work: slack app → aws lambda → claude → workato, u just have to build and host the lambda + bolt on retries / observability urself. or skip both and have workato itself listen to slack and call claude via their anthropic connector — workable if workato is already the heavy lift in ur stack.
for what u described i’d just go Slack Trigger → AI Agent node (Tools Agent mode, claude as the model, snowflake + SF MCP tools attached) → HTTP Request to workato. the Tools Agent mode is the closest thing to ur “agent sits in claude with instructions” mental model — claude picks which tools to call and in what order, n8n just hosts the trigger and the outbound HTTP.
one thing worth knowing — running claude’s tool-calling through n8n means each tool call is an n8n node execution, not a claude-internal call. tradeoff is better observability (u see every step in the exec log) but slightly slower than if claude could run the loop natively.
I think at this point it is more compliacted. Claude has now agents (in the console and it has integrations, Slack is an example right away) and Claude Workspace, which can act as an agent and perform all the tasks you mention. But the advantages of n8n are still valid in terms of observability and also to be more independent of a provider. Right now I am hesistant to use any agent frameworks of Anthropic or OpneAI or Google directly, since I want to be the owner of the business logic and to be able to just change the LLM, if the next better one comes around next week So, there are always more ways to complete something but the n8n way keeps you more independent and flexible in terms of LLM usage and workflow management.
So, there are always more ways to complete something but the n8n way keeps you more independent and flexible in terms of LLM usage and workflow management.