← All episodes · Peter Yang
I Tested Gemini Spark: What Google's AI Agent Can Actually Do in 21 Minutes
Overview
Peter Yang does a live, use-case-by-use-case walkthrough of Gemini Spark, which he calls the most beginner-friendly personal agent on the market. His framing: the single biggest step most people can take toward being AI-native is moving from a chatbot that answers questions to an agent that does work in their apps — and by his numbers, while ~60% of the world's 8 billion people use a chatbot, only 0.04% use a real agent. With Gemini's 900 million monthly active users, Spark could be the bridge. (He discloses the video is sponsored by Google but gives his honest take throughout.)
He demos five workflows live: Gmail inbox triage scheduled as a daily 7am briefing, podcast interview prep that researches guests and writes question docs, Tokyo vacation planning across Google Flights/Maps/Docs with price-drop monitoring, a Bay Area summer camp tracker in Sheets with registration alerts, and Spark's skills system — including a 'match my writing style' skill that builds a voice card from your Google Docs and a 'get more perspectives' council-of-experts skill. His verdict: Spark's differentiator is deep default integration with Google products (no API setup), its big gap is zero third-party integrations, and its real destiny is replacing the Gemini chatbot for mainstream users.
Topics
From chatbot to agent: why Spark matters
Peter's thesis is adoption math. Roughly 60% of 8 billion people use a chatbot like ChatGPT, but only 0.04% use a real AI agent — the gap his opening chart makes vivid. Tools like Hermes and Claude Code are powerful but intimidating for beginners, while Gemini already has 900 million monthly active users, so Spark (a toggle inside Gemini, currently paid-subscribers-only) can be the on-ramp: same familiar surface, but an agent that acts in your Google apps instead of just answering questions.
The recurring beginner-friendliness point: to give other agents access to his Gmail he had to wrestle with Google Cloud Console, connect APIs, and download JSON credentials; Spark works with all Google products out of the box.
Demo 1: Gmail triage as a scheduled morning briefing
The prompt: scan the last 7 days of Gmail and produce a three-section morning triage — urgent emails, emails to unsubscribe from, and a recap — each with sender, one-line summary, why it matters, and a direct Gmail link. Spark classified his real inbox live: urgent items included summer camp registration, an email from his accountant, and podcast scheduling with OpenAI; unsubscribe candidates included a local news newsletter and ads.
The follow-up is the actual payoff: 'Schedule this as a briefing every morning at 7:00 a.m. PST' turns the one-off prompt into a recurring automation delivered in chat. Peter says he has reached the point where he relies on his agent to triage his inbox rather than reading every email himself.
Demo 2: Podcast prep from Google Calendar
The task: every weekday morning, check Google Calendar for podcast-interview-looking events in the next 7 days, research each guest via public YouTube interviews and articles, and write a 'podcast prep' Google Doc of suggested questions. For his two next-day interviews, Spark took about five minutes and produced a doc per guest — the Tariq (Anthropic) doc linked his actual GitHub, organized themes like 'HTML is the new markdown' and human-agent interaction, and drafted questions per theme; the Ritesh doc included a table of demos to ask about.
When the questions ran too long for his taste, he just told Spark to shorten them in the doc, and it edited the Google Doc in place. His framing: it's like hiring a podcast-prep employee — you still check its work, but once the template is right it saves an hour of manual research per week.
Demo 3: Vacation planning across Flights, Maps, and Docs
Asked for the cheapest 12-day round trip to Tokyo in December, Spark used its native Google Flights integration — notable because Flights has no public API, so this is genuinely hard for other agents — and cross-referenced his December calendar and his existing vacation-planning Google Doc before suggesting dates (early-to-mid December beats his originally planned Dec 16–28 on price). A follow-up about Shinjuku hotels switched it to Google Maps, surfacing family apartment-hotels including one he'd actually stayed at.
Pushing further, he had it critique his existing vacation doc: it recommended a multi-city ticket (SFO→Tokyo, return from Fukuoka), flagged that mountain stops on his itinerary are high-altitude with late-December snow risk, then updated the doc on request. Finale: 'monitor this itinerary and alert me when SFO→Tokyo drops below $800' — a standing price watch. His take: a combination of flights, maps, web search, and docs that works like a travel agent.
Demo 4: Summer camp tracker in Sheets
Instead of the 'crazy spreadsheets' Bay Area parents pass around, he asked Spark to build a detailed Google Sheet of the best July summer camps with sign-up links, prices, descriptions, age ranges, and registration-opening dates. The result impressed him twice over: the camps were near where he lives rather than generic Bay Area picks (Spark's memory at work), and spot-checking a sign-up link showed it actually resolved to a live 'register now' page.
He scheduled an alert for when a specific camp opens registration, with the honest caveat that this only works if the camp publishes that information — 'AI helps you save time, but you still have to use your critical thinking.'
Demo 5: Skills — voice cards and a council of experts
Skills are reusable saved prompts with descriptions, trigger conditions, and onboarding steps — a schedules view alongside them shows every automation created during the video (inbox triage, podcast brief, flight monitor, camp alert), each editable, runnable on demand, or deletable. The 'match my writing style' skill scanned his Google Docs and produced a voice card: direct and active, candid and down-to-earth, polite but decisive, short punchy sentences, highly structured — 'pretty far removed from typical AI slop copy.' With the style guide saved, he can paste any draft and have it edited to match his voice.
The 'get more perspectives' skill answers a question through personas (operator, skeptic, visionary). Asked about body recomposition, it cited his actual squat progression and body-fat target from his emails and health doc — against which he contrasts generic Gemini's boilerplate ('progressive overload, high protein') to make the episode's core point: the agent's value is the personal context it can reach.
Verdict: differentiator, gap, and destiny
The number-one differentiator is deep default Google integration — Gmail, Calendar, Flights, Maps, Docs, Sheets — with none of the API plumbing other personal agents require, plus the standard agent toolkit of scheduled tasks, default and custom skills, and output into the apps you already use. The big missing piece: no third-party integrations at all (he can't connect his bank), leaving Spark less powerful than other personal agents for now, though he expects Google to fix that.
His conclusion: because most people already live in Google products, Spark is probably the most beginner-friendly personal agent to set up if you have a Gemini subscription — and if it succeeds, its real role is to replace the Gemini chatbot across 900 million users, turning them from question-askers into people whose agent does work in their favorite apps.
Visuals
| Gemini Spark | Other personal agents | |
|---|---|---|
| Google apps access | Built-in: Gmail, Calendar, Flights, Maps, Docs, Sheets | Manual API setup via Cloud Console + JSON credentials |
| Google Flights | Native (no public API exists) | Workarounds at best |
| Third-party integrations | None yet — can't connect your bank | Broad connector ecosystems |
| Scheduling & skills | Scheduled tasks, default + custom skills | Comparable |
| Best for | Beginners already living in Google apps | Power users needing reach beyond Google |
Takeaways
- The AI-native leap is chatbot → agent: ~60% of people use a chatbot, only 0.04% use an agent that actually does work in their apps.
- The pattern in every demo: run a task once, verify the output, then schedule it — inbox triage, podcast prep, price watches, and registration alerts all become standing automations.
- Spark's moat is zero-setup Google integration (including Google Flights, which has no public API); its current ceiling is zero third-party integrations.
- Agent output needs a template-tuning pass: check the work, correct it conversationally ('make the questions shorter'), and the corrected version becomes the recurring standard.
- Personal context is the product: the same question answered generically by Gemini and personally by Spark (from your email, docs, calendar) is the whole difference.
- A 'voice card' built from your own docs is a practical guard against AI-slop writing — save it as a skill and edit drafts against it.
Notable quotes
The number one thing that most people can do to become more AI native is to switch from using a chatbot that can answer questions to an AI agent that can actually do work for you in your favorite apps.— Peter Yang
For me to use Hermes or some other agent to actually look at my Gmail, I had to go to Google Cloud Console and connect all these APIs and download a JSON... but Spark works with all Google products right out of the box.— Peter Yang
It's almost like hiring a podcast prep employee to do this work for you automatically.— Peter Yang
AI helps you save time, but you still have to use your critical thinking and just make sure that things actually work.— Peter Yang
If this product succeeds, it's to basically just replace the Gemini chatbot across Gemini's 900 million users.— Peter Yang