AI agents are finally starting to feel less like clever demos and more like coworkers. With Manus 1.6 and its new Max agent, tasks that used to stall, break, or need constant hand-holding are starting to complete themselves, end to end.
At the same time, Nvidia is quietly locking in the plumbing that all of this runs on, from cluster schedulers to open models built for agents, not just chat. Put together, these moves say a lot about where AI is heading next: more autonomy, more reliability, and much bigger ambition.
This breakdown walks through what changed in Manus 1.6, what Nvidia just bought and released, and why it matters if you care about agents doing real work, not just answering questions.
Manus 1.6: Closing the Gap to Real AI Agent Autonomy
The central idea behind Manus 1.6 is simple: finishing a task once for a demo is easy, finishing it correctly and autonomously over and over is hard.
This release is aimed squarely at that gap. Manus rebuilt its core agent architecture so it can plan better, reason better, and stay stable through long, messy workflows with less human help.
At a high level, Manus 1.6 introduces three big changes:
- Manus 1.6 Max: the new flagship agent, built for planning and problem solving
- Full mobile development support: agents can now build mobile apps, not just web tools
- Design View: an interactive visual workspace for image creation and editing
According to the team, one of the clearest improvements is oneshot task success. More tasks now reach the finish line on the first try, without humans stepping in to fix broken steps or rewrite prompts.
That improvement comes from deeper reasoning and stronger internal planning, not just faster execution.
In double-blind tests focused on user experience, Manus measured a 19.2% jump in reported satisfaction. Testers rated three things in particular:
- The quality of the outputs
- The accuracy of the results
- How reliably the agent used tools without breaking workflows
If you compare this release to what open agents are starting to do elsewhere, it fits into a broader shift toward more autonomous systems, similar to what you see with the open source AI agent that surprised OpenAI and Google.
For a deeper technical overview straight from the source, the Manus team breaks down the release in their own post on Introducing Manus 1.6: Max Performance, Mobile Dev, and Design View.
Inside Manus 1.6 Max: The Flagship Agent for Dependable Work
The center of gravity in this update is Manus 1.6 Max, now the default choice if you care more about dependability than quick demos.
Better planning and problem solving
Max runs on a more advanced internal architecture that is tuned for long-horizon planning and problem solving. In practice, that shows up as:
- Fewer interruptions, fewer corrections
- Fewer half-finished runs that stall at step 7 of 12
- More tasks that complete cleanly without you jumping in mid-way
The spotlight feature is still oneshot task success. You hand the agent a full workflow, not just a single question, and it is much more likely to:
- Break the problem into steps
- Call the right tools in the right order
- Keep track of what has already been done
- Finish without you having to restart or rewrite the instructions
That shift changes how you use an agent. Instead of hovering over it like a nervous parent, you can treat it more like a junior teammate that mostly takes care of itself.
Upgrades in key workflows
Max does not only think better in the abstract. Several concrete workflows get meaningful upgrades.
Wide research
Manus has a feature called wide research that spins up multiple sub-agents in parallel to explore different angles of a topic. With 1.6, all of those sub-agents now run on the Max architecture, not a weaker variant.
That leads to:
- Deeper analysis on each branch
- Better cross-checking between sub-agents
- Fewer “weak links” that drag down the final summary
So if you ask for a full market landscape or a multipage research brief, every branch of that tree is using the strongest reasoning the system offers.
Spreadsheet-heavy work
Max is also much stronger on spreadsheets and data:
- Complex financial models
- Advanced data analysis workflows
- Automated report creation from large datasets
These are the types of tasks where fragile formulas and small errors used to pile up. With Max, multi-step calculations are more stable and less likely to collapse if one sheet or range changes.
If you are interested in how similar agentic systems are already affecting real financial workflows, the breakdown on AI agents transforming finance workflows gives a useful parallel.
Web development that cares about UX
On the web side, Manus 1.6 goes beyond simply generating working code.
Max pays more attention to:
- UI aesthetics
- Layout structure
- Interactive behavior
That matters because internal tools live or die on usability, not just whether the code compiles.
With the new stack, you can describe things like:
- An internal app that parses uploaded invoices
- Editable forms that update underlying data
- Summary statistics dashboards
and get interfaces that feel closer to a real product than a weekend hack.
Competitor analysis with real structure
Max also helps create detailed feature comparison matrices. These are not just narrow side-by-side tables with three bullet points each.
You get structured, multi-dimensional comparisons that capture:
- Product features
- Pricing models
- Implementation tradeoffs
- Strategic pros and cons
That kind of structured view is much more useful when you are deciding what to build next or which product to buy.
Benchmarks and real-world wins
Inside Manus, agents are tested on scenarios that look like real use, not abstract multiple-choice exams. The internal benchmarks show gains across nearly every category, with the biggest jumps in complex multi-step tasks that need both accuracy and sustained reasoning.
Here is a simple view of where Max moves the needle the most:
| Area | What Improved With Manus 1.6 Max |
|---|---|
| Complex multi-step workflows | Holds context better and finishes more tasks in a single run |
| Spreadsheet-heavy tasks | More stable formulas and fewer silent failures in long calculations |
| Research and analysis | Stronger parallel branches and better cross-checking |
| Web app building | Cleaner layouts and more usable internal tools |
| Tool use and orchestration | Fewer broken tool calls and more reliable flows |
The real outcome is straightforward: tasks that used to take several attempts now often finish successfully in one shot. That saves time, reduces friction, and makes the agent feel more like a collaborator than a demo toy.
Expanding Beyond Web: Manus Mobile Development Support
Until now, many AI agent platforms have lived almost entirely in the browser. Manus 1.6 changes that by adding full mobile development support.
You describe the app you want, and the agent handles the end-to-end build:
- Screen layouts
- Navigation structure
- Core logic
- API wiring
For a lot of products, mobile apps are often the primary interface. Being locked to web-only tools meant you could prototype dashboards but not deliver a complete experience.
With mobile support in place:
- Startups can sketch, test, and ship cross-platform ideas much faster
- Internal tools can reach field teams, not just people at laptops
- Experiments do not have to stop at a proof-of-concept web panel
Combined with the upgraded web stack, Manus now covers a much larger slice of what real products need to look useful in daily work.
Design View: Visual Control Without Fighting Prompts
Design View is the third big addition in Manus 1.6, and it tackles a very common pain: trying to do precise visual work with only text prompts.
Instead of typing “move the logo slightly left and make the subtitle smaller” ten times in a row, you get an interactive canvas where you can:
- Click to make local edits to specific parts of an image
- Add or modify text inside images with high-quality rendering
- Combine several images into more complex compositions
Under the hood, Manus still uses modern image generation models. The canvas simply gives you fine-grained control that feels closer to traditional design tools.
That combination makes it much easier to iterate on:
- Marketing visuals
- Product mockups
- Presentation graphics
without getting stuck in the “prompt, wait, reject, re-prompt” loop.
If you are interested in how other companies are thinking about richer visual and multimodal agents over the next few years, the breakdown on Google's vision for 2026 omnimodel AI is a helpful companion.
Availability and Limited-Time Deal
Manus 1.6 is live for all users right now. When you start a new task, you can choose which agent to use, including Max.
During the rollout, there is also a pricing twist: the Max agent runs at a 50% reduced credit cost for a limited time. That discount is meant to lower the barrier to trying Max in real workflows instead of only small tests.
Limited-time promo: If you have a backlog of tasks you have been nervous to hand off, this is a good period to see how far a more autonomous agent can go.
For full details and official notes, Manus shares everything in their own Manus 1.6 Max release post.
Nvidia’s Big Moves: Securing AI Infrastructure and Agent Models
While Manus pushes agents closer to real autonomy at the task level, Nvidia is busy securing what those agents actually run on: large compute clusters and open models tuned for agentic use.
Two announcements stand out:
- The acquisition of SchedMD, the company behind Slurm
- The release of the Nvidia Nemotron 3 family of open models
Acquiring SchedMD: Owning Slurm for Massive AI Scale
Nvidia announced that it is acquiring SchedMD, the team behind Slurm, one of the most widely used open-source workload managers in high performance computing and AI.
A few key points:
- Slurm has been around since 2002
- SchedMD was founded in 2010 by Morris Jette and Danny Auble, the lead developers behind Slurm
- Danny Auble is the current CEO of SchedMD
If you have ever worked around large clusters, Slurm is the system that:
- Queues jobs
- Allocates GPUs and CPUs
- Coordinates massive workloads across thousands of machines
Nvidia has worked with SchedMD for more than a decade and calls Slurm critical infrastructure for generative AI. In its public statements, Nvidia says SchedMD will continue running Slurm as open-source, vendor-neutral software, while Nvidia invests more to speed up integrations with different systems.
The company shared more details in its official post on the Nvidia announcement on acquiring SchedMD and Slurm.
The message is clear: GPUs matter, but orchestration matters just as much once you scale. By bringing Slurm closer, Nvidia tightens its grip on the full stack that large AI workloads depend on.
Nemotron 3: Open Models Built for Agents, Not Just Chat
Alongside the acquisition, Nvidia introduced Nemotron 3, a family of open models described as some of the most efficient options for building accurate AI agents.
These models are designed for agentic systems from the start, not only single-turn chat. The family includes three variants:
- Nemotron 3 Nano
A small, efficient model for targeted tasks where speed and cost dominate. - Nemotron 3 Super
A mid-tier model aimed at multi-agent setups where several AI systems coordinate or interact. - Nemotron 3 Ultra
The largest model, built for complex tasks that need deeper reasoning and broader capabilities.
In the launch materials, Nvidia CEO Jensen Huang repeated a theme he has pressed for a while: open innovation is foundational for AI progress. Nemotron is framed as a way to give developers an open, transparent platform for building agentic systems at scale.
You can find the full technical and product details in Nvidia’s announcement on the Nemotron 3 family of open models.
Open Models and Physical AI
Nemotron 3 is not Nvidia’s only recent open move. In parallel, the company has:
- Announced Alpameo R1, an open reasoning vision-language model for autonomous driving research
- Added more workflows and guides for its Cosmos world models, which are open source under a permissive license and aimed at physical AI systems
If you connect those pieces, the pattern is pretty clear. Nvidia wants to be the default provider for:
- Robotics
- Autonomous vehicles
- Embodied systems in real environments
All of those need:
- Huge amounts of compute
- Robust scheduling on clusters
- Efficient models that can run reliably in the wild
Other players are pushing in similar directions with their own agent stacks, as seen in systems like China's Kimi K2 multimodal AI breakthrough. Nvidia’s angle is to supply the infrastructure and the open models that everyone else builds on top of.
My Experience and What I Learned From Agentic Workflows
Watching tools like Manus Max evolve, and comparing them with systems such as Abacus AI Deep Agent autonomous browser automation, has changed how I think about “using AI”.
Early on, I treated AI as a smarter search box: ask a question, get an answer, move on. Over time, the real value has shifted toward handing over whole workflows, not just prompts.
Here is what stands out for me:
- Planning matters more than raw IQ
The biggest difference between a basic model and a strong agent is not vocabulary, it is planning. When an agent can pause, outline steps, and decide what to do next, the final result feels much closer to human work. - Oneshot success changes your habits
When you trust that a task will likely complete in a single run, you start queuing up bigger jobs. You think in projects, not prompts. - Visual feedback is underrated
Features like Design View (more on that in a second) highlight how helpful it is to see what the agent is doing, not just read logs. It is similar to watching a teammate share their screen while they work. - Autonomy exposes weak instructions
The more autonomous agents become, the more they expose vague thinking on the human side. If your brief is fuzzy, the output will be too. Clear goals and constraints matter a lot more once the agent actually acts on them.
Overall, the main thing I have learned is this: the biggest unlock is not a single smarter model, it is systems that can think, act, and recover from small mistakes inside a real workflow.
Final Thoughts: Where AI Agents Go From Here
Manus 1.6 and Nvidia’s latest moves live at different layers of the stack, but they point in the same direction.
On one side, you have agents like Max that are better at planning, finishing tasks in one run, and handling real workflows across web, mobile, spreadsheets, and design. On the other, you have infrastructure and open models tuned for agentic systems at scale, from Slurm-managed clusters to Nemotron 3.
The common thread is autonomy. Less babysitting. More workflows handed off. More systems that can plan, act, and correct themselves inside real work.
If you are building or adopting AI agents today, the question is not just “Which model is smartest?” It is “Which systems can actually carry a task across the finish line without constant supervision?”
That is where the action is heading next.
0 Comments