Comet, Cursor, and the CB Insights Lens I Use Before I Build

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

Before I write agent code I want a market lens and a research agent that can fan out. CB Insights for the premium landscape, Perplexity Computer and Comet for parallel browse-and-build, Cursor when the answer finally becomes a repo. That order has saved me from the most expensive demo of all: a polished agent for a market that already solved the pain with a CRM checkbox and a bored champion.

Parallel research is the product

Perplexity Computer running four jobs in parallel is the habit I copy on every stack: fan out questions, merge into a table, only then decide. The stock app built without traditional coding is a proof of path — not a portfolio piece I would sell unreviewed to a regulated client.

Comet in that walkthrough is the browser-shaped coworker. I watch its permissions the way I watch Browser Operator: no password fields, no payment pages, no admin consoles without me in the chair with coffee and a kill key.

CB Insights as a pre-build filter

Premium market data will not invent my wedge. It will stop me from rebuilding a crowded category with a thinner distribution plan and a prettier landing page. I extract who already owns the workflow, where budgets sit, and which "AI feature" is a thin wrapper over last year's dashboard.

If the lens says the buyer already has three vendors, I need a sharper job-to-be-done or I walk. Walking early is a billable skill. Building anyway is a personality flaw I am trying to retire.

Long agents need long tests

Cursor's ten-hour agents, Copilot autonomous mode, Claude Code memory and voice, Nano Banana 2 for visuals — the toolchain is noisy on purpose. I pick Cursor when acceptance tests live in the repo. I pick research agents when the acceptance test is "did we learn enough to say no without embarrassing ourselves."

Ten-hour agents need ten hours of tests and a kill switch. Unbounded autonomy is how cloud bills become horror stories and how half-finished branches become tribal lore.

Vendor-switch headlines as dependency risk

Large buyers switching AI vendors are useful as dependency stories, useless as morality plays. I care whether my client's data residency and audit story still holds if a vendor gets dropped next quarter.

Model choice stays swappable behind my own eval harness. If swapping means rewriting the product, I did not build a product — I built a hostage situation with nice UI.

A research week I reuse

Monday: landscape notes into a one-page brief. Tuesday–Wednesday: parallel jobs filling a comparison table with sources. Thursday: kill or commit — Cursor spikes only on commit. Friday: threat model for the commit path, including who can press stop.

Skipping Monday is how I fall in love with a tool. Skipping Thursday is how I build anyway. The table is how I say no without sounding lazy in a room that rewards motion.

What "built a stock app" taught me

No-code and agent-built demos compress the path from idea to UI. They do not compress fiduciary duty, market microstructure knowledge, or the part where a wrong number costs someone money. I keep demos labeled as demos in the README and in the sales call.

If a client wants the demo promoted, we add data contracts, monitoring, and a human who can halt trading-shaped actions. Until then it is a teaching artifact, not a product.

Comet's permission hygiene

Browser agents see too much by default. I create a clean profile, seed only the tabs it needs, and revoke access when the research window closes. Leaving Comet logged into everything is how "research" becomes "incident."

I also ban screenshotting of secrets into the research table. Tables get redacted. Raw captures stay local and get deleted after the brief is written.

The sentence I end pitches with

We will spend a week learning whether to build. If the answer is no, you still get the table and the lens. That sentence filters clients who only wanted a build theatre.

Saying no with the table in the room

I print or project the comparison table when I recommend walking away. Abstract caution sounds like fear. A table with three vendors owning the workflow sounds like judgement. Comet and Cursor can wait in the hallway until the table has spoken.