I Stopped Asking Chatbots and Started Hiring Agents

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

The mental model that finally clicked for me: Manus is not a smarter chatbot. It behaves like a junior employee who executes. You give it a job, it opens tools, recovers from errors, and brings back an artifact. Advice is optional. Delivery is the point.

I worked through six concrete use cases from my notes. Each one changed how I brief agents on client work.

Shopping without handing over the wallet

In the Amazon shopping test, Manus researched products, compared options, and recovered when it hit an error page. The human still entered the OTP and payment. That split is the whole security lesson. Let the agent navigate and assemble. Keep authentication and money in human hands.

When I adapt this for work, the agent may draft a vendor shortlist and fill a comparison table. I am the one who clicks “purchase” or signs the MSA.

Cleaning a downloads mess it broke itself

On a downloads cleanup, Manus wrote a Python script, watched the script fail, rewrote it, and sorted twenty-three PDFs into categories. Self-correction matters more than a perfect first draft. I now ask agents to log each repair step so I can see whether they are looping or learning.

For my own machine I still run cleanup inside a disposable folder copy first. Agents that reorganize files can also delete the wrong tree if you point them at home directory.

Ads analysis that asks before it acts

Across eleven Meta Ads accounts, the agent asked permission first, then recommended what to pause or scale and listed the top three next moves. Permission-first is the pattern I want in every production agent. Silent “optimizations” on ad spend are how budgets disappear.

My version of this workflow: read-only access for analysis, a written recommendation, and a separate human step to apply changes in Ads Manager.

Slack archaeology and parallel research

One run searched more than two hundred Slack channels. Another researched two hundred Indian fitness influencers in parallel and returned a structured table. Parallelism is the unlock. A chatbot answers one thread. An agent can fan out, then merge.

I use the same shape for competitor scans: define columns up front (audience, posting cadence, offer, proof), let the agent fill rows, then I score the table myself.

Shipping a tiny app, not a mockup

Manus built a React Native onboarding flow, produced an Android APK, and gave an Expo Go QR preview for iPhone. That is past “here is some JSX.” It is a build you can tap.

I would still run the APK through my own checklist — permissions, hardcoded secrets, debug endpoints — before any client sees it. Speed without a security pass is just technical debt with a QR code.

Telegram as the remote control

Telegram integration lets you delegate with a voice note and get the report on your phone. That is how I want field updates to work when I am away from the desk: speak the task, receive the artifact, approve the risky step later.

How I brief an agent now

I write three blocks before any Manus-style run: goal artifact (what file I want back), tools allowed (browser, Slack read-only, local folder X), and hard stops (no send, no pay, no delete). If I cannot fill those blocks, I am not ready to hire the agent yet.

For the influencer-research shape, I paste the column schema in the brief itself. For the ads shape, I require a permission prompt before any mutate call. For the app-build shape, I demand a local preview path and a list of permissions the APK requests.

Browser Operator sits underneath several of these flows. It turns a normal browser into hands for the agent. That is exactly why I keep financial and identity pages off-limits unless I am watching the session.

The hard rule I will not relax

The notes are explicit: keep human approval for purchases, messages, deletions, and publishing. I treat that as non-negotiable. Browser Operator-style agents that can click anything are powerful and dangerous in the same breath.

If you only remember one thing from this post, remember the employee metaphor — and remember that even good employees do not get the company card without asking.