Day 7: Building 'Jerry' — Master Agent, Sub-Agents as Tools, the List-Then-Act Pattern, and Calling Any API from n8n

A 'message a model' AI node is stateless and hard-wired to one provider ('you either choose an OpenAI or a Gemini'). An AI Agent node keeps conversational memory across turns and lets you swap the LLM without rebuilding. AI nodes remain the right choice for single-task jobs such as the Day 3 spam filter; agents are for conversations ('reply to the same person with x' needs memory), multi-tool coordination, and any natural-language command - a raw Gmail node cannot interpret 'send an email to abc'.

Planning precedes both: 'you can never create a workflow without a plan' - the first NODE is a trigger, the first STEP is enumerating commands per domain and the compound commands that span them.

Akhil walks through the alternatives and knocks each down: sequential chaining (every agent runs on every request - credits wasted), parallel branches (same waste), Switch/IF on keywords (cannot parse natural language), an AI-node classifier feeding a Switch (fails on compound cross-domain commands). 'That calls for something called a master agent and sub agent architecture.' n8n's recent update added an 'AI Agent' TOOL node type distinct from the action node; only the tool type can be nested - 'an action node can never be a tool.'

Build: the Email Agent is converted into a tool under a new Master Agent; the sub-agent's input is 'automatically defined by the model' (the master decides what to pass). Master memory 10, sub-agent memory 5 - the master must hold every domain. Master system prompt: route by domain to Email/Calendar/Meeting agent tools; for combinations 'use the appropriate tools in the order of requirement.' Live test: 'send an email ... as well as create an event titled after party' - both happened. Calendar and Meeting agents are then 'Lego pieces': duplicate chat model + memory, swap the tools.

Every n8n app tool is configured as a Resource (message, draft, event, calendar) and an Operation (Get Many, Send, Reply, Delete, Create, Update, Get Availability). If the operation you need is not listed, check 'a level above' - Draft is its own resource; Get Availability lives under Calendar, not Event. Tools are Lego: duplicate and change resource/operation to mint a new skill.

Reply, delete, update - and Fireflies' fetch-transcript - all need an ID you do not possess. The fix is never hardcoded: the system prompt says 'first read via Get Many, extract the ID, then use the appropriate tool.' Without a system message 'my agent does not have a direction' and will hallucinate tool use - 'every agent is as good as the instruction that you give.' (Expression mode is used for prompts only because it gives a bigger editor.) Live debugging: wrong Gmail account produced 'I'm not authorized'; the default email signature/attribution was disabled in Send/Reply options.

Models are trained to a cutoff and, ungrounded, treat that as today, so relative dates in calendar commands land in the wrong year. Fix: attach n8n's built-in Date and Time tool (no configuration) and write into the Calendar Agent's prompt 'before you perform any action, get the current date.' The workflow time zone - three-dot menu, workflow settings - governs what 'now' means. The master does not need the date; email carries its own timestamps.

Memory is a cost dial, not a free feature: Simple Memory's context-window length is how many turns get re-sent with every message, so 'if my input size is large... you'll have to pay more money.' Hence 5 for sub-agents, 10 for the master, and 'no' to a learner asking for 500.

For any SaaS without a native node: register and copy its API key; ask the vendor's documentation AI (or ChatGPT/Perplexity) in plain English how to do the thing - Fireflies needs two calls, list transcripts (to get IDs) then fetch transcript + attendees by ID, the same list-then-act shape; copy the cURL and Import cURL into an HTTP Request Tool node, which fills headers, body and placeholders. Paste the key yourself: 'although all the APIs or any other LLMs will give you a structure, they don't have your API key.' Because tool fields cannot reference agent-extracted variables, the second call's JSON body (with the transcript-ID slot) is constructed by the agent at runtime, instructed in the system prompt. Live: 'give me the summary of the last meeting' chained both calls and returned the summary.