Twelve months ago, "autonomous agent" was vapourware. Twelve weeks ago, it was demo-ware. Today, four platforms have shipped products that are productive for real work, with real cost structures, and real differentiated sweet spots. If you are deciding which one to build on (or use), this is the honest read after a month of running tasks through each.
One note before we start: Google discontinued Project Mariner on May 4, 2026[2][3]. Its technology was folded into a new Chrome feature called Auto Browse, gated behind Google AI Pro and AI Ultra. References below to "Mariner" should be read as "the agent capabilities now living inside Chrome Auto Browse".
_Disclosure: my Manus referral link gives both of us bonus credits if you want to try it after reading this. I have tested all four; Manus is the one I would put my own money on for most tasks, and I am up front about that bias._
The four-way head-to-head
| Spec | Manus | OpenAI Operator | Claude Agent SDK | Chrome Auto Browse |
|---|---|---|---|---|
| Architecture | Multi-agent | Single agent | Developer SDK | Single agent |
| Browser control | Yes | Yes | Via tools | Yes (Chrome native) |
| Sub-agents / parallel | ✅ Wide Research | ❌ | ✅ (you build) | ❌ |
| App / website builder | ✅ | ❌ | ❌ | ❌ |
| Payments integration | ✅ Stripe | ❌ | ❌ | ❌ |
| Desktop access | ✅ | ❌ | ✅ (Computer Use) | ❌ |
| Pricing model | Per-credit | ChatGPT Pro | API per-token | Bundled AI Pro / Ultra |
| Best for | Research, app prototypes | Routine browser tasks | Custom agents | Chrome users |
Let me walk through where each one wins.
Manus — best for research and full-stack prototypes
Manus 1.6 Max with Wide Research is the most capable general-purpose agent platform right now[1]. It is the only one of the four that:
- Runs multiple parallel sub-agents on a single goal
- Builds full apps end-to-end with database + Stripe + SEO baked in
- Has a desktop app with local file access
- Generates polished slide decks, financial models, and research reports as deliverables
Where Manus loses: pricing is per-credit, which makes cost-prediction harder. And the platform is opinionated — you do not have fine-grained control over how the agent thinks.
Best for: research-heavy work, market analysis, prototyping apps, content generation at scale.
OpenAI Operator — best for routine browser tasks
Operator is the simplest. A single agent, in a browser, doing a task. Book a flight. Order groceries. Fill in a form. It is fast, reliable for short tasks, and bundled into ChatGPT Pro at no marginal cost if you already pay for that.
Where Operator loses: anything that needs more than one agent, anything that needs to touch your local filesystem, anything that needs to build an artefact (an app, a doc, a report).
Best for: routine errands. The use case it ships for, it nails. Nothing more.
Claude Agent SDK — best when you need control
Claude Agent SDK is Anthropic's developer toolkit, not a consumer product. You build the agent. It runs on Claude Opus 4.7. You decide what tools the agent has, what data it sees, what guardrails apply.
Where Claude Agent SDK loses: there is no UI. There is no "give it a goal" entry point. You write code.
Where it wins: you have complete control. If you are building an agent that handles regulated workflows, sensitive data, or custom integrations, the SDK is the right starting point. Claude Opus 4.7's long-context consistency makes hour-plus agent runs reliable in a way 4.5 was not.
Best for: production agents in your own product. Custom internal tooling. Anything where the agent needs to do exactly what you say and only what you say.
Chrome Auto Browse — best for Chrome power users
Chrome Auto Browse is the Mariner replacement, native in Chrome on AI Pro and Ultra. Sits in your browser. Watches what you do. Offers to do things for you.
Where Auto Browse loses: it only lives in Chrome. It cannot escape the browser. It does not run parallel agents. It does not build artefacts.
Where it wins: tight Chrome integration. If your work happens almost entirely in browser tabs, Auto Browse removes friction in ways the others cannot match.
Best for: power users whose work is browser-shaped. Researchers, recruiters, ops people.
How to choose
Three questions to ask yourself:
- What does the deliverable look like?
- Research report → Manus Wide Research - Filled form / completed errand → OpenAI Operator - Custom agent for your own product → Claude Agent SDK - Browser-tab workflow you do every day → Chrome Auto Browse
- How much control do you need?
- Maximum control, your own infrastructure → Claude Agent SDK - Some control, fast iteration → Manus - No control, just want it done → Operator or Auto Browse
- What is your cost structure?
- Variable per-task → Manus credits or API tokens - Fixed monthly → Operator (ChatGPT Pro) or Auto Browse (Google AI Pro/Ultra)
The honest verdict
If I had to pick one as a generalist, it is Manus. The combination of Wide Research, the app builder, and the Stripe / SEO integrations covers more ground than the others combined. The architecture (multi-agent + sub-agents) is the right answer for non-trivial tasks. And the per-credit pricing makes it transparent what you are paying for.
But "generalist" is doing a lot of work in that sentence. Each of the four wins clearly in its sweet spot. The right answer is to use Manus for research and prototypes, Operator or Auto Browse for routine browser stuff, and Claude Agent SDK when you are building production agent features into your own product. They are not really competing — they are filling different slots.
Which is why I run all four, depending on the task. If you want to test Manus first (it is the one most worth a real evaluation), my referral link gives you bonus credits.