Briefing · 04/06/2026

From tasks to governed entities

Gemini Spark, ChatGPT Agent, Microsoft Copilot agents and Meta Business Agent point to the same shift: AI work is moving from isolated prompts to projects, persistent workers and governed digital entities.

The first wave of consumer AI products trained people to think in prompts.

A user asked for an answer, a summary, an image, a draft, a table, or a piece of code. The unit of work was the task. It began with an instruction and ended with an output.

That phase still matters, but it no longer explains where the market is moving.

The more important shift is from task execution to persistent delegated work. Google’s Gemini Spark page describes a personal AI agent that can work in the background around the clock, operate across connected Google apps, run tasks, use skills, follow schedules, and check with the user before major actions.

Google’s own product post makes the direction clearer. It describes Gemini Spark as a 24/7 personal AI agent that can keep working in the background after the laptop is closed or the phone is locked. Google also says Spark will expand through connected apps, MCP connections, custom sub-agents, and local browser operation.

That is not simply another interface for a prompt.

It points toward an evolved operating model for AI work. The system can hold state, monitor progress, use connected services, and continue after the initial instruction has been given. The product question moves from “can the model answer?” to “what role does this system play inside the work environment?”

The clearer ladder is:

Task -> Project -> Persistent Worker -> Governed Entity

Task

A task is a bounded instruction.

Summarise this document. Draft this reply. Generate this image. Create this chart. Extract these rows. Rewrite this section. Compare these options. Turn this meeting transcript into actions.

The task layer is now broadly commoditised. Frontier systems from OpenAI, Google, Anthropic, Microsoft and Meta can all perform a large class of task-level work well enough for everyday use.

That does not make task execution unimportant. It makes it a baseline capability.

Standalone products whose only durable claim is generic task completion now sit in the exposed part of the market. If a tool exists mainly to generate a blog post, summarise a PDF, draft a reply, create a basic image, or turn notes into a first-pass deck, it has to compete with bundled AI inside the platforms where users already work.

The undefined middle of the market is not every specialist product. It is the product that has no durable context, no specific workflow, no production control, no delivery channel, and no authority model.

Project

A project is a stateful container.

It has context, artefacts, versions, goals, constraints and accumulated decisions. It may include documents, files, source material, collaborators, a deadline, a budget, a preferred style, a compliance requirement, a source list, or a record of previous choices.

This layer matters because most valuable work is not a single prompt. It is a sequence of related actions inside a durable context.

The project layer explains why the market is moving beyond chat. Users do not only need a model to answer a question. They need a system that understands what the current work is, what has already been decided, what material is in scope, what style or policy applies, and what the next useful step should be.

This is why project-style surfaces have become strategically important. ChatGPT Projects, Claude Projects, Notion workspaces, Microsoft 365 files and Teams context, Google Drive and Workspace context, Canva brand kits, Zapier interfaces, and similar environments all point in the same direction.

The work is becoming organised around persistent context.

A project is not yet a worker. It is the container in which a worker can become useful.

Persistent Worker

A persistent worker is assigned to continue activity over time.

It can monitor, wait, return, update, compare, alert, draft, escalate or continue. Its value comes from continuity rather than one-off intelligence.

This is where the current platform shift becomes more significant.

OpenAI’s ChatGPT Agent is designed for interactive workflows where the user can interrupt, steer, take over the browser, and use connectors. OpenAI also describes recurring tasks and warns that agentic systems introduce new risks because they can work with user data, access connectors, and act on websites where the user is logged in.

Microsoft Copilot Studio is described as a SaaS agent platform for building AI agents, agentic workflows and multi-agent processes. That is a different product category from a chat assistant. It is infrastructure for delegated work inside business processes.

Meta Business Agent shows the same pattern through customer channels. Meta says its agent can answer business-specific questions, recommend products, book appointments, qualify leads, decide when a team member should step in, and take action through connected business systems. Meta is also building a platform for businesses to customize and deploy these agents at scale across WhatsApp, Messenger, Instagram and Meta Business Suite.

These products differ in surface, buyer and operating environment. The shared direction is more important than the product differences.

They are not competing only to answer questions. They are competing to become the delegated worker inside a user’s digital environment.

Governed Entity

The final layer is the governed entity.

A persistent worker becomes operationally significant when it has identity, memory, permissions and continuing authority. It may act through a user account, a service account, a browser session, a workspace, an app integration, a channel, or a device.

It may read files, draft messages, update systems, notify people, trigger workflows, book appointments, create artefacts, or share information with third parties.

At that point the central question is no longer “how capable is the model?”

The central question becomes: who is this entity acting for, where is it allowed to operate, what can it remember, and what controls apply when it acts?

This is where much of the current AI product discussion remains underdeveloped. The market has plenty of language for models, prompts, benchmarks and agents. It has less settled language for the authority boundary around a persistent digital worker.

That boundary includes practical questions:

Which identity is the worker using?
What data can it access?
What actions can it perform without approval?
What events require human confirmation?
What logs are retained?
Can a user reconstruct what happened?
Can an administrator suspend or constrain it?
Can personal and work contexts remain separate?
What happens when instructions conflict?
What recovery path exists after a failed or unsafe action?

These questions are not secondary. They define whether persistent AI work can move from impressive demonstration to reliable operation.

The Platform Direction

The platform companies understand this.

Google has distribution through Gmail, Docs, Sheets, Slides, Drive, Calendar, Chrome, Android and the Gemini app. Gemini Spark makes that distribution more active by placing a persistent worker inside the Google context.

Microsoft has distribution through Microsoft 365, Teams, Windows, Edge, Copilot Studio, Power Platform, Entra identity, Intune policy and enterprise administration.

Meta has distribution through WhatsApp, Instagram, Messenger, Facebook, Meta Business Suite and the business conversations already happening across those channels.

OpenAI has built a broad assistant surface with apps, connectors, files, browsing, analysis, agent mode, recurring tasks and user supervision controls.

Anthropic is moving Claude into high-context projects, integrations, tool use and computer-use workflows.

The competitive direction is clear. The largest platforms want to own four layers at once:

the place where work is requested;
the context in which work is organised;
the worker that continues after the request is made;
the controls around that worker’s authority.

That does not mean every specialist product disappears.

It means the undefined middle of the market becomes exposed. A product that wraps a model around a generic task will struggle. A product that owns a high-control workflow, a trusted delivery channel, a specific production environment, or a governance layer can still matter.

The evolved market will not be divided cleanly between “AI platforms” and “AI tools”. It will be divided by operational role.

Some systems will answer tasks. Some will organise projects. Some will act as persistent workers. Some will become governed entities within a broader operating environment.

A Better Evaluation Frame

This distinction gives buyers a more useful way to evaluate AI products.

Instead of asking whether a product “has AI”, ask what role it plays.

Is it producing an output?

Is it maintaining a project context?

Is it continuing work over time?

Is it allowed to act with authority?

Can that authority be governed?

Those questions reveal the real product category.

A task tool should be judged by quality, speed and cost.

A project system should be judged by context, continuity, collaboration and artefact management.

A persistent worker should be judged by reliability, monitoring, escalation, supervision and recovery.

A governed entity should be judged by identity, permissions, logging, auditability, data boundaries, approval paths and containment.

This is also the more serious frame for OpenClaw and similar systems.

The opportunity is not to build another general assistant. The opportunity is to manage the authority boundary around delegated AI work: identity, permissions, logs, approvals, memory, local access, browser use, recovery and separation between personal and organisational contexts.

The next stage of AI adoption will be defined less by prompt quality and more by entity control.

The task layer is becoming normal. The project layer is becoming the workspace. The persistent worker is becoming the new operational actor. The governed entity is the layer that will decide whether that actor can be trusted.

Sources

Was this useful?

Quick signal helps Rob sharpen future briefings.