What OpenAI’s GPT-6 Astra, Google’s Project Astra and Meta’s Muse Glimmer tell us about where AI is going

For years, one of the easiest ways to measure progress in artificial intelligence was to ask a simple question:

How well can the machine answer us?

That question is becoming less useful.

The next phase of AI is increasingly about what happens after the answer.

Can the system research something, make decisions, use software, operate tools, create an output and complete a task?

In other words:

Can we give AI an objective rather than an instruction?

That is the shift toward what I would call the Agent Era.

And three developments in 2026 make that shift particularly interesting: OpenAI’s GPT-6 Astra, Google’s Project Astra and Meta’s Muse Glimmer.

They are not identical technologies. In fact, they represent three very different ideas about what the next generation of AI could become.

1. OpenAI: AI that can do the work

OpenAI’s GPT-6 Astra represents perhaps the clearest move from conversation to execution.

OpenAI describes Astra as its most capable model for reasoning, coding, computer use, research and professional work. It can operate computers and browsers, work through multi-step tasks and produce documents, spreadsheets and presentations while following templates and changing requirements.

The difference is subtle but enormous.

The old interaction was:

“Tell me how to create a marketing plan.”

The emerging interaction is:

“Research this market, develop the marketing plan, analyse the competitors and prepare the presentation.”

The AI is no longer simply giving you information.

It is becoming capable of participating in the workflow.

That is a profound change for knowledge work.

2. Google: AI that understands your world

Google’s Project Astra takes a different path.

Rather than focusing primarily on controlling your digital desktop, Astra is built around perception.

It is Google’s research effort toward a universal AI assistant capable of understanding real-time visual and audio information and interacting naturally with people. Google is already moving some of the capabilities explored through Project Astra into Gemini Live, Search and emerging hardware such as glasses.

Imagine pointing your phone at something and asking:

“What am I looking at?”

But then going much further.

The AI can understand what you are seeing, hear what you are saying, remember context and respond conversationally.

This points toward a world where the AI interface isn’t necessarily a chatbot window.

It could be the camera.

The microphone.

Your glasses.

Your surroundings.

The AI becomes less like an application you open and more like an assistant that can perceive the environment around you.

3. Meta: AI that can run locally

Meta’s Muse Glimmer introduces another important direction.

Instead of putting all the intelligence in the cloud, what happens if increasingly capable AI can run directly on your own computer?

Muse Glimmer is a 30-billion-parameter open-weight model released under an Apache 2.0 licence. Meta says it is optimised for local agent workflows and can run on a Mac or PC with a single consumer GPU. It supports use cases including local agents, function calling and coding.

This matters enormously.

Local AI means there are circumstances where your data doesn’t need to leave your machine.

It can also reduce dependence on constant internet connectivity and cloud infrastructure.

For developers, businesses and researchers, that opens another possibility:

AI that you control.

Three directions. One major shift.

These aren’t mutually exclusive.

In fact, I suspect they are pieces of the same much bigger picture.

From chatbots to agents

This is where the terminology matters.

A chatbot waits for you to ask something.

An agent can potentially take a goal, break it into steps, use tools, observe what happens, adjust its approach and continue until the task is completed.

That changes the economics of software.

Think about a marketing team.

Today, a campaign might involve:

Researcher → Strategist → Copywriter → Designer → Video editor → Media buyer → Analyst.

Tomorrow, one person might coordinate several AI agents performing parts of that workflow.

The human doesn’t disappear.

The role changes.

Instead of doing every individual task, the human increasingly becomes the person defining the objective, setting constraints, reviewing decisions and managing the system.

And this is where it gets interesting for Africa

For businesses in Nairobi, Lagos, Accra or Johannesburg, the implications are significant.

AI agents could allow very small teams to perform work that previously required much larger departments.

A five-person company could potentially conduct market research, produce creative assets, analyse data, build presentations, manage customer communication and support marketing campaigns with AI assisting across the workflow.

But local AI could be equally important.

Connectivity, cloud costs, data sovereignty and privacy are real considerations.

A model that can run on hardware you control introduces a different economic proposition.

This is why Muse Glimmer is interesting beyond the technical community.

It is part of a broader question:

Who gets to access AI, and where does that intelligence live?

But there is a catch

The moment AI moves from generating content to taking action, the risks change.

A chatbot producing a bad paragraph is annoying.

An agent making a bad decision inside your CRM, sending an email to the wrong customer, changing a database or interacting with a vulnerable system is something else entirely.

OpenAI has explicitly said GPT-6 Astra reaches its “Critical” cybersecurity capability threshold under its Preparedness Framework. The company says that, with appropriate tools and access, Astra can identify previously unknown vulnerabilities and develop exploits across well-protected systems without someone guiding every step.

This means the Agent Era isn’t simply a conversation about smarter models.

It is a conversation about:

Permissions.

Identity.

Access.

Monitoring.

Accountability.

And ultimately:

Trust.

If an AI can act on your behalf, we need to know exactly what it is allowed to do.

Where does this leave humans?

This is the question I find most interesting.

Google is exploring AI that can see and understand the world around us.

OpenAI is building AI that can increasingly operate the digital world.

Meta is pushing toward powerful agentic models that can run locally.

And these approaches may eventually converge.

Imagine an AI that can:

See your environment

↓

Understand your objective

↓

Reason through the problem

↓

Use your computer

↓

Access authorised tools

↓

Complete the task

↓

Report back to you

At that point, the computer stops being merely something we operate.

It becomes something that can operate on our behalf.

And that brings us back to the question that may define the next decade of AI.

For years, we asked:

What can AI answer?

Now we need to ask something much more consequential:

What are we willing to let AI do?

That is where the Agent Era really begins.