C7 EST | 14 Day AI Sprint·Day 11 | Office Hours + Agents with Google Workspace Studio and Lyzr ·4:21:00

Day 11: Two Alternatives to n8n — Google Workspace Studio (VIP-Email Demo) and Lyzr Agent Studio (Knowledge Bases, Sub-Agents, a Manager, and API Deployment)

Akhil Guest mentor (Days 3, 7) - demos Google Workspace Studio and builds the e-commerce multi-agent system in Lyzr · Harshad Office-hours co-mentor - confirmed to lead the Day 14 monetization session · Samriddhi Office-hours co-mentor - discovery-call-first consulting methodology · Uthappa Host - Phase 2 wrap-up logistics, hackathon announcement

▶LMS lesson

The short version

Google Workspace Studio is n8n-lite inside Gmail: Starter = trigger, Actions = nodes, check/filter = if/else, plus AI skills and an 'Ask Gemini' step. Demo: an email containing 'immediately' -> Decide node scores urgency against written criteria -> Ask Gemini extracts TLDR/summary/ask -> VIP label -> Google Chat ping. Needs a registered Workspace business domain; personal Gmail will not unlock it. 'Very new... first 0.1% of folks' seeing it; n8n 'already covers exhaustively.'
Lyzr is 'a chat-first agent system that we can deploy in about 5 minutes': create a Knowledge Base (upload PDFs, test retrieval by typing 'shoes'), create an agent (Name, Model, Role, Goal, Instructions - let 'Generate with AI' draft them and recommend tools), attach a tool (Hacker News -> Top Stories), set memory (5-10 messages), Create.
Modularity is the keyword: three scoped sub-agents (Shopping Assistant on the catalog, Refund Specialist on the policy with an empathetic tone, Returns Specialist on the tickets) under a Manager enabled by a toggle, because one agent's context grows with the catalog and demands 'a bigger brain' for everything. A combined 'buy shoes and return my umbrella' request was routed to both.
Deployment: Launch Agent -> name/category/industry -> hosted chat UI. The old HTML embed seems gone; integration is now by API - a vibe-coded front end posts the message and shows the reply in '10-15 lines of code', key in the environment file; Studio edits propagate live.
Verdict: Lyzr = chat-agent scope with RAG and orchestration abstracted and model costs bundled (Pro ~$99, Startup ~$30-40 as heard); n8n = any node, any media, granular control, but you manage RAG, deployment and API keys yourself. 'Modularity' and 'trade-off' are the words.
Office hour: protect workflow IP by selling outcomes on retainer, not JSON; discovery call -> whiteboard the flow -> approval -> build; enterprise clients come from referrals and free pilots; a 'thin wrapper' is a prompt on someone else's model, a 'thick wrapper' adds proprietary data and workflow.

The concepts

Google Workspace Studio: n8n vocabulary inside Gmail (and its hard limits)

how-to

If everything a client does lives in Google, the automation tool can live inside Gmail too - as long as they own a Workspace domain.

Access: a registered Google Workspace business account (akhil@outskill.com); personal Gmail shows no 'Workspace Studio' icon. Vocabulary maps one-to-one to n8n: Starters (on schedule, email received, chat message, Sheet row added/updated, Drive file added, meeting, form response); Actions (Tasks, Docs, Drive, Sheets add/update row, Gmail label/star) that are 'compatible and super fast within the Google ecosystem'; check/filter branches; AI skills - recap 100 emails, extract information, summarize, decide - and Ask Gemini for anything else. Turning a workflow on 'is as good as publishing inside my n8n canvas.' A discovery page offers templates (daily AI-news briefing 'Monday 8 AM, weekdays').

Limits stated: Google-only (no Outlook), very early ('first 0.1% of folks... getting the demo'), and a fallback rather than a replacement because n8n already covers the ground.

Do it in this order

01Log in with a Workspace business account; open the Workspace Studio icon in Gmail
02Starter: 'on email received'
03Decide node: paste urgency criteria (deadlines, production/financial risk, senior-leadership request = urgent; routine = not)
04True branch -> Ask Gemini: extract action items as TLDR + summary + ask
05Gmail action: add label 'VIP'
06Google Chat action: send the formatted summary
07Turn the workflow on; send yourself an email containing 'immediately' to test

Why it matters

For Paul's Google-Workspace clients this is a zero-install way to ship small automations without another vendor.

Go deeper

In one line: Workspace Studio = Starter -> Actions/AI skills -> branches inside Gmail; requires a Workspace domain; G-Suite-only; early stage.

Requires a registered Workspace business domain; personal Gmail won't unlock it (l3186065 0:11)

Starter = trigger; Actions = nodes; check/filter = if/else; AI skills + Ask Gemini (l3186065 0:20-0:22)

Demo: 'immediately' email -> Decide -> Ask Gemini -> VIP label -> Google Chat (l3186065 0:13-0:28)

'On' = published, like an n8n canvas (l3186065 0:18)

G-Suite only; 'very new'; n8n covers it exhaustively (l3186065 0:31-0:32)

▶ Watch this taught:

Lyzr: knowledge base, one agent, one tool - in five minutes

how-to

Type 'shoes' into the knowledge base before you build anything. If the right chunks come back, the agent will work; if not, no prompt will save it.

Platform: 'reimagined agentic AI automation', chat-first, compliance badges (HIPAA, CCPA, ISO, GDPR, SOC 2); provider-agnostic models (OpenAI incl. cheap tiers, Anthropic Claude 3.5/3.7 Sonnet, Google, Bedrock, Perplexity, Kimi K2, Llama); memory either Lyzr-native or AWS Agent Core in your own account; a voice feature 'recently introduced... use with a pinch of salt.' Knowledge Base: Create new -> name/description -> default vector store -> add text / website / file (drag a PDF) -> query it directly to preview the closest chunks. Agent: Create Agent -> Name, Description, Model (GPT-4o Mini for the demo) -> Role, Goal, Instructions; 'Generate with AI' interviews you, asks about output format, and recommends tools (Perplexity for breadth vs a News API for control) because live information needs tool calling 'just the way we have done it in n8n.' Tools -> Add -> Configure more tools -> Hacker News -> Top Stories -> Connect (then edit instructions to name the tool). Memory slider (5, 10), optional file/image output, and a Hallucination Manager gated to higher plans. Observed flaw: the news briefer returned a 2023-dated item when asked for the last 24 hours - the knowledge base is static, so 'dynamic' data needs a real tool.

Do it in this order

01Knowledge Base -> Create new -> name + description -> default vector store -> Create
02Add File (PDF) / Add Website / Add Text; run a test query to inspect retrieved chunks
03Home -> Create Agent -> Name, Description, Model
04Generate with AI: describe the agent; answer its questions; 'Use this prompt' to fill Role/Goal/Instructions
05Tools -> Add -> Configure more tools -> pick connector -> Select Action -> Connect; mention the tool in Instructions
06Toggle Knowledge Base, set Memory (5-10), Create; test in chat

Why it matters

It is the fastest route from a PDF to a working RAG chatbot in this KB - and its limits (static KB, plan-gated guardrails) are the ones to price in.

This idea elsewherecontrastDay 4 exposes every RAG parameter; Lyzr hides them behind a default vector store

Go deeper

In one line: Knowledge Base (upload + test retrieval) -> Agent (Role/Goal/Instructions via Generate with AI) -> Tool (native connectors) -> memory -> Create.

'Chat-first system... deploy in about 5 minutes'; HIPAA/CCPA/ISO/GDPR/SOC 2 (l3186065 0:41-0:42)

Models from OpenAI, Anthropic, Google, Bedrock, Perplexity, Kimi K2, Llama; memory native or AWS Agent Core (l3186065 0:48, 1:25)

Knowledge Base create + in-place retrieval test (l3186065 0:44-0:46)

Agent fields drafted by Generate with AI incl. tool recommendation (l3186065 0:47-0:53)

Tool attach: Hacker News Top Stories (l3186065 0:54-0:55)

Hallucination Manager gated to higher tiers; 2023-dated result observed (l3186065 0:58, 1:02)

▶ Watch this taught:

Modularity is the keyword: three specialists and a manager in Lyzr

One agent that knows the catalog, the policy and every ticket needs a bigger brain for every question. Three small agents and a manager do not.

The e-commerce build on three documents (product catalog for 'Awesome Shopper' in INR/USD, return & refund policy, active return tickets): Shopping Assistant (catalog KB, 5 chunks, basic retrieval, memory 10), Refund Specialist (policy KB, instructed tone - 'oh, that's strange', 'sorry to hear that'), Returns Specialist (tickets KB). The argument for splitting: a single agent's context grows with the catalog and ticket volume and drags the whole scope onto an expensive model; modules let you extend 'without having a bumped price for the entirety of operations.' Manager: a normal agent with 'Make this manager' toggled below Instructions, which reveals 'Add an agent' to attach sub-agents; Lyzr recommends high-reasoning models for it (GPT-4o or Claude 3.5); no knowledge base of its own, deeper memory allowed; a canvas view under orchestration shows the graph with editable connection descriptions. Test: 'I want to purchase new shoes and return my umbrella' hit both specialists, including a ticket lookup by customer name.

Why it matters

The third time this sprint the manager/sub-agent shape appears (Day 7 n8n, Day 10 principle, Day 11 Lyzr) - it is the durable architecture.

This idea elsewhereparallelSame architecture as Jerry, in a hosted platform appliesGoldilocks scope realized as separate specialists

Go deeper

In one line: Scoped sub-agents each with their own KB and tone, under a manager agent (toggle) on a high-reasoning model; canvas shows the routing graph.

Split to avoid context growth and 'bigger brain' cost across the whole scope (l3186065 1:05-1:07)

Sub-agents: Shopping (catalog), Refund (policy, empathetic tone), Returns (tickets) (l3186065 1:08-1:14)

'Make this manager' toggle -> 'Add an agent' (l3186065 1:15-1:16)

Manager model: GPT-4o or Claude 3.5; no own KB; deeper memory (l3186065 1:16-1:17)

Canvas/orchestration view with editable connection descriptions (l3186065 1:22-1:23)

Combined shoes + umbrella request routed to both (l3186065 1:19-1:21)

▶ Watch this taught:

Deploying a Lyzr agent by API, and the Lyzr-vs-n8n verdict

The embed button he remembered is gone. What is left is better for a builder: an API and fifteen lines of front end.

Launch Agent -> confirm quality -> Name, Category (productivity / customer experience), Industry (retail), Function (customer service), optional welcome message -> Launch App (private by default) -> View Agent opens a hosted chat UI. No HTML embed could be found live; Lyzr appears to have moved to API integration: a custom front end (vibe-coded) posts the message field to the deployed agent endpoint and renders the response - '10, 15 lines of code' - with the key in the environment file as taught in the vibe-coding session; Studio changes propagate without redeploying.

Verdict on 'if starting today': Lyzr is chat-agent scoped - no first-class video/image or broad workflow nodes, limited connectors (HubSpot, Notion, 'major things'), but RAG, vector stores and orchestration are abstracted and model costs bundled into the subscription (Pro ~$99, Startup ~$30-40, as heard). n8n gives 'granular control plus more automation possibilities' at the price of managing RAG, deployment and API keys yourself. Whether Jerry could live on Lyzr depends on whether its connectors cover Gmail, calendar and meetings.

Why it matters

A clean statement of the abstraction trade-off Paul faces with every hosted agent platform.

This idea elsewhereparallelCatalyst Basecamp 4 bridges a front end to n8n the same way

Go deeper

In one line: Lyzr deploys to a hosted chat or an API for custom front ends; pick Lyzr for fast chat agents with bundled costs, n8n for breadth and control.

Launch Agent -> metadata -> Launch App -> View Agent (l3186065 1:26-1:27)

HTML embed not found; API integration is the path (l3186065 1:28-1:31)

Front end posts message via API, renders reply; ~10-15 lines; key in env file (l3186065 1:32-1:36)

Studio edits propagate live (l3186065 1:35-1:36)

Lyzr = chat scope, limited connectors, abstracted RAG, bundled pricing (Pro ~$99, Startup ~$30-40) (l3186065 1:37-1:42)

n8n = any node, granular control, self-managed RAG/keys (l3186065 1:38-1:40)

▶ Watch this taught:

Office Hour 7 distilled: protecting workflow IP, discovery-first consulting, thin vs thick wrappers

'How do I get enterprise clients?' Referrals and free pilots. There is no shortcut without a portfolio.

Commercial: keep workflows confidential by selling outcomes on retainer rather than handing over JSON; monetize a custom-GPT coaching bot as a metered SaaS via the OpenAI Assistants API and Stripe (also for revenue shares); one website for a multi-service agency is fine; Samriddhi's method - discovery call -> whiteboard the flow -> client approval -> build; enterprise entry is referrals and free pilots. 'AI wrapper': thin = a prompt on someone else's model; thick = proprietary data + workflow on top (Uber and Perplexity as thick-wrapper examples). Technical: client-side Gmail auth via Firebase rather than credentials in nodes; Google Cloud Console API + custom webhooks to consolidate calendars from non-API sources; Outlook node has near-parity with Gmail; free model routes - Kimi K2, Ollama GPT-OSS, Groq credits (no free API trials from the big labs); AWS EC2 or DigitalOcean for your own server; Perplexity Comet or Claude's Chrome extension for browser automation like LinkedIn requests (HeyReach for rate-limit safety); Apify debugging - swap the static 'dataset' URL segment for the variable, unpin data, enable community nodes, rotating-IP nodes for Meta blocks; exact image matching = reference image + Google Vision API + an accuracy layer; small-doc RAG for coding tools = file injection over Pinecone; Lindy is a simpler prompt-driven alternative to n8n.

Why it matters

The discovery-first method and the thin/thick wrapper distinction are the two ideas Paul can reuse in the next client conversation.

Go deeper

In one line: Retainer over JSON; discovery -> whiteboard -> approve -> build; referrals/pilots for enterprise; thick wrappers add data + workflow; practical routes for free models, own servers, browser automation.

Protect IP: sell the outcome on retainer, not the JSON (l3319327 0:10-0:12)

Discovery call -> whiteboard -> approval -> build (l3319327 0:25-0:26)

Enterprise via referrals and free pilots (l3319327 0:24-0:25)

Thin wrapper = prompt on a model; thick = proprietary data + workflow (l3319327 0:44-0:45)

Free model routes: Kimi K2, Ollama GPT-OSS, Groq; no lab free API trials (l3319327 0:16, 0:27)

Firebase for client OAuth; Cloud Console API + webhooks for non-API calendars (l3319327 0:06, 0:18)

Apify fixes: dataset variable, unpin, community nodes, rotating IPs (l3319327 0:32-0:34, 0:39-0:43)

Image matching: reference image + Google Vision API + accuracy check (l3319327 0:45-0:46)

▶ Watch this taught:

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01Google Workspace Studio: n8n vocabulary inside Gmail (and its hard limits)Workspace Studio = Starter -> Actions/AI skills -> branches inside Gmail;›

Workspace Studio = Starter -> Actions/AI skills -> branches inside Gmail; requires a Workspace domain; G-Suite-only; early stage.

Requires a registered Workspace business domain; personal Gmail won't unlock it (l3186065 0:11)

Starter = trigger; Actions = nodes; check/filter = if/else; AI skills + Ask Gemini (l3186065 0:20-0:22)

Demo: 'immediately' email -> Decide -> Ask Gemini -> VIP label -> Google Chat (l3186065 0:13-0:28)

'On' = published, like an n8n canvas (l3186065 0:18)

G-Suite only; 'very new'; n8n covers it exhaustively (l3186065 0:31-0:32)

02Lyzr: knowledge base, one agent, one tool - in five minutesKnowledge Base (upload + test retrieval) -> Agent (Role/Goal/Instructions via Generate with AI) -> Tool (na…›

Knowledge Base (upload + test retrieval) -> Agent (Role/Goal/Instructions via Generate with AI) -> Tool (native connectors) -> memory -> Create.

'Chat-first system... deploy in about 5 minutes'; HIPAA/CCPA/ISO/GDPR/SOC 2 (l3186065 0:41-0:42)

Models from OpenAI, Anthropic, Google, Bedrock, Perplexity, Kimi K2, Llama; memory native or AWS Agent Core (l3186065 0:48, 1:25)

Knowledge Base create + in-place retrieval test (l3186065 0:44-0:46)

Agent fields drafted by Generate with AI incl. tool recommendation (l3186065 0:47-0:53)

Tool attach: Hacker News Top Stories (l3186065 0:54-0:55)

Hallucination Manager gated to higher tiers; 2023-dated result observed (l3186065 0:58, 1:02)

03Modularity is the keyword: three specialists and a manager in LyzrScoped sub-agents each with their own KB and tone, under a manager agent (toggle) on a high-reasoning model;›

Scoped sub-agents each with their own KB and tone, under a manager agent (toggle) on a high-reasoning model; canvas shows the routing graph.

Split to avoid context growth and 'bigger brain' cost across the whole scope (l3186065 1:05-1:07)

Sub-agents: Shopping (catalog), Refund (policy, empathetic tone), Returns (tickets) (l3186065 1:08-1:14)

'Make this manager' toggle -> 'Add an agent' (l3186065 1:15-1:16)

Manager model: GPT-4o or Claude 3.5; no own KB; deeper memory (l3186065 1:16-1:17)

Canvas/orchestration view with editable connection descriptions (l3186065 1:22-1:23)

Combined shoes + umbrella request routed to both (l3186065 1:19-1:21)

04Deploying a Lyzr agent by API, and the Lyzr-vs-n8n verdictLyzr deploys to a hosted chat or an API for custom front ends;›

Lyzr deploys to a hosted chat or an API for custom front ends; pick Lyzr for fast chat agents with bundled costs, n8n for breadth and control.

Launch Agent -> metadata -> Launch App -> View Agent (l3186065 1:26-1:27)

HTML embed not found; API integration is the path (l3186065 1:28-1:31)

Front end posts message via API, renders reply; ~10-15 lines; key in env file (l3186065 1:32-1:36)

Studio edits propagate live (l3186065 1:35-1:36)

Lyzr = chat scope, limited connectors, abstracted RAG, bundled pricing (Pro ~$99, Startup ~$30-40) (l3186065 1:37-1:42)

n8n = any node, granular control, self-managed RAG/keys (l3186065 1:38-1:40)

05Office Hour 7 distilled: protecting workflow IP, discovery-first consulting, thin vs thick wrappersRetainer over JSON;›

Retainer over JSON; discovery -> whiteboard -> approve -> build; referrals/pilots for enterprise; thick wrappers add data + workflow; practical routes for free models, own servers, browser automation.