Engineering notes / Muhammad Kashif
watsonx and Manus My Computer: Enterprise Agents on a Tight Leash
IBM's message at a Gartner Mumbai stage — better data beats better models, with HR, procurement, and sales already running through agents — landed next to Manus My Computer, a desktop agent that organises files, builds a…
watsonx and Manus My Computer: Enterprise Agents on a Tight Leash
By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan
IBM's message at a Gartner Mumbai stage — better data beats better models, with HR, procurement, and sales already running through agents — landed next to Manus My Computer, a desktop agent that organises files, builds apps, controls software, and taps local GPU. Enterprise appetite and desktop power in one week. My leash got shorter on purpose: seller agents that update CRM and auto-draft RFPs only after the data layer is trustworthy and the scopes are dull.
Data quality before model theatre
If CRM opportunities are messy, a seller agent will create messy RFPs faster. IBM's framing matched what I see in Pakistan enterprise pilots: the bottleneck is definitions, deduplication, and ownership, not another leaderboard model. I price data cleanup as a prerequisite, not an optional phase-two fantasy.
Agents amplify the warehouse you have. They do not invent a clean warehouse out of heroic prompting.
Seller agents on a rail
Updating CRM, creating opportunities, drafting proposals so humans spend time with clients — excellent when drafts stay drafts until a human sends. I wire stage changes that need manager approval, and I ban agents from marking deals closed won without a human.
Auto-create is allowed. Auto-commit to the customer is not. The difference is a button and an audit id.
Manus My Computer with folder diets
Desktop control across Mac and Windows, local GPU for local models, building desktop apps — power tools for ops teams. I start Manus-shaped agents on a diet: one project folder, no password manager, no browser profiles with bank cookies, no email send without confirmation.
Local AI is attractive for privacy. Local AI with access to everything is just a breach without needing the cloud.
HR and procurement agents
Enterprise stories about HR and procurement agents only work with clear policy encodings. I ask who maintains the policy when labour law or vendor rules change. An agent with a stale policy is a confident liability.
Procurement agents that can issue POs get hard money caps and dual control. Curiosity is not a spending authority.
Enterprise leash checklist
Classification labels on every connector. Retention set. Human roles named for revoke, approve, and incident. Eval sets from real messy tickets, not demo scripts. Quarterly access review. If any line is missing, the pilot stays in a business unit sandbox.
Gartner-stage confidence does not replace the checklist. It motivates funding it.
What I refuse to automate first
Final customer emails. Salary changes. Vendor bank detail edits. Legal hold content. Those wait until the boring automations — drafting, routing, summarising — are measured and dull.
Leaders often want the dramatic use case first. Dramatic first is how pilots die publicly.
Pakistan delivery note
Many local firms want the IBM slide without the data programme. I split proposals: phase A data contracts and CRM hygiene, phase B seller-assist agents, phase C desktop agents for internal ops. Skipping A is a polite decline from me.
Leash on my own machine
Even solo, Manus-like tools get a scratch user account. Enterprise habits start at home. If I will not trust an agent with my own files unbounded, I will not sell that trust to a client.
Phase-A gate in the SOW
No seller-agent phase B until phase A data contracts are marked done by a named data owner. IBM-style slides go in the appendix. Manus desktop toys stay on scratch accounts until B is earned. The leash is contractual, not vibes.
Dual control on POs
Procurement agents draft purchase orders; a second human must approve above a small threshold. Manus may organise the PDF folder. It may not be the second human. Watsonx-style enterprise stories inherit the same dual control in my SOWs.
Scratch user on my Mac
Manus-shaped tools run as a separate OS user with a short home directory. Enterprise leash habits start on my own machine so I do not sell a fantasy I will not live. Watsonx pilots inherit the same separation of duties story.
Seller agent cannot close-won
CRM stage "closed won" is human-only. Draft RFPs yes. Celebrate in Slack yes. Mark won no. IBM-inspired automation still respects who owns the number that hits the board forecast.