The $100 agent is now $20

Welcome to the Toolkit, the Friday half of The AI Debrief. Tuesday is what happened. Friday is what to do about it: one tool worth your time, one play you can run today, and nothing you have to buy.

Here's what we're getting into this week. The agent that runs your inbox while you sleep just dropped to $20 a month, the catch is who it's allowed to work for, and the fix for both is a page you write in twenty minutes.

Gemini Spark just got 80% cheaper

Google added Spark to the $19.99 AI Pro plan in the US. It launched in May as an Ultra-only feature starting at $99.99 a month, so the cheapest way in just dropped by 80%.

What it actually is: three pieces. Tasks, where you hand it something multi-step and walk away. Schedules, where you set it to fire on a clock or a trigger. And Skills, where you teach it how you do a thing once so you stop re-explaining it every time. It works across Gmail, Calendar, Docs, Sheets, Slides and Drive, and it keeps running when your laptop is shut.

The examples Google leads with are the boring, correct ones. Scan the inbox every morning and summarize what matters. Pull meeting notes into a formatted report. Draft a reply when a specific person emails. Build a cited digest on a topic you follow.

The piece worth your attention is Skills, because it's the closest thing anyone has shipped to a written job description for an agent. Google's own example: read the last 50 emails I wrote, turn that into a style guide for how I write, save it as a skill called ghostwriter, and call it every time I ask for a draft. That isn't a prompt. That's a documented process the agent reloads on every run. Hold that thought, because it's the whole second half of this email.

So if your business runs on Google Workspace, the agent that automates email cannot touch the inbox where your actual work lives. You'd be pointing it at a personal Gmail. The cost barrier came down and the permissions barrier stayed exactly where it was.

There's a second read here, and it's the one I'd act on. Google just put an always-on agent at the same price as a chat subscription. That isn't a discount, it's a category move: background work is now the default thing you're paying for, and every other lab has to answer it. Which makes the useful question this week not "should I subscribe." It's which of my recurring work is specified well enough to hand over - because once the agent can only see half your accounts, the constraint stops being the tool and starts being you.

Try it here →

Write the job description first

Every agent I've handed real work to has failed for the same reason, and it was never the model. I knew the task cold, I'd done it four hundred times, and I had never once written it down. So I'd type half an instruction, get back something almost right, fix it by hand, and quietly decide the tool wasn't there yet.

The tool was fine. The spec was missing.

Two things make or break this.

Decision rules, not steps. "Check email" is a step, and it's useless - the agent already knows how to open an inbox. "Archive it if it's a newsletter I haven't opened in a month, flag it if a client's name is in the subject line, leave everything else alone" is a rule. Rules are the part that lives in your head and nowhere else, which is exactly why the agent can't guess them and exactly why writing them down is the whole job.

Stop conditions. Name what it must never do without checking with you. Send anything to a client. Spend money. Delete instead of archive. Reply to anyone outside a list you write out. This is the line between an agent you trust with real work and one you supervise so closely that you'd have been faster doing it yourself.

Twenty minutes, and you walk away with one page. That page is the asset, not the subscription. Tools change every six weeks - the story above is a tool changing price by 80% in ten weeks - but a written process outlives all of them, and it's portable. The same page works in Spark, in Cowork, in whatever ships next quarter. And if you do end up in Spark, Skills is literally a place to store it.

That's the difference between an agent that saves you an hour a day and one you quietly turn off on day three.

5 SaaS tools you can cancel now

▶ Video · 10:41 · free in the AI Resource Vault

Calendly and Notion AI are two of the five. Calendly runs $12 to $25 a month, Notion AI is a $10 add-on on top of what you already pay Notion, so that's roughly $22 to $35 a month between those two alone.

It isn't a cancel-everything pitch. For each of the five the guide lays out what replaces it, the exact setup, and where the replacement falls down - Calendly's public booking link is a real thing you give up, and if you're a coach taking 30 inbound bookings a week you should keep paying for it. Direction, not dogma.

The other three are in the video.

Watch it in the vault →

Do the job description one before Monday. Twenty minutes, costs nothing, and worst case you end up with a written version of something you've been carrying in your head for years - which is worth having whether or not you ever hand it to an agent.

That's the Toolkit.

See you Tuesday for the Debrief.

From Drew and The AI Debrief team

P.S. What's the one task you'd hand an agent tomorrow if you actually trusted it? Hit reply and tell us. We read every one, and the good answers turn into future issues.

P.P.S. New here, or skipped one? Every past edition is in the archive.