Skip to content

Agents · · 8 min read

Always-On Agents Have Arrived. The Interesting Part Is the Permission Model.

In one month OpenAI released GPT-6 Astra and then Sol and Luna, Google shipped Gemini 3.8 Flash and Gemini 4 Argon, Anthropic released Claude Opus 5.5 and Sonnet 5.5, and the list prices came down with them — Opus 5.5 launched 20% below its predecessor, and OpenAI’s Luna at ten cents per million input tokens. But the release that matters most to how businesses will work was not a model. At DevDay on 29 September OpenAI announced dots: always-on agents inside ChatGPT, each with its own cloud computer and browser. Four days earlier Microsoft unveiled Autopilot, a persistent Copilot agent given a role and a goal. Agents that keep working when you close the tab are now a product category.

From chat to colleague

Until now most AI at work has been conversational: you ask, it answers, the exchange ends. An always-on agent is different in kind. It holds a standing role, watches for work, acts across several steps and tools, and comes back to you with results or questions. Dots run on OpenAI’s flagship model with their own browser and machine; Autopilot is defined by a role and a goal and keeps going until the goal is met.

For a business this is the shape that automation has been reaching for: not a chatbot bolted onto a website, but a worker that can be given a lane. The capability has been close for a year. What changed in September is that it arrived in the tools companies already pay for.

Every serious agent ships with a policy

Look at what the vendors built alongside the agents. Dots come with Custom Rules that let administrators allow, block or require approval for each kind of action, and enterprise admins must switch them on before anyone can use them. Autopilot is scoped by its role. Across the industry the same three buckets appear: things the agent may do on its own, things it must ask about first, and things it may not do at all.

That is not a safety afterthought; it is the product. An agent without a permission model is a liability with a login. The companies shipping these systems have concluded that the approval gate is where trust is earned, and they have put it in front of administrators rather than hiding it in a settings page.

What this means for your own systems

When a client asks us to automate approvals, claims, scheduling or document review, we now design the workflow in exactly those three buckets. Routine, reversible steps — classifying, drafting, routing, summarising — run on their own. Anything that touches money, access, legal records or a customer’s inbox stops at a named person with the agent’s proposed action and reasoning laid out. A short list of actions is simply not available to the agent at all.

The audit log is the other half. Each action, who or what took it, and what the agent was told at the time should be recorded as a matter of course. When an always-on worker has been running for a month, the question “why did it do that?” must have an answer.

Start with a lane, not a department

The temptation with a persistent agent is to hand it everything. The teams getting value are doing the opposite: one well-defined lane — chasing overdue documents, triaging inbound requests, preparing a daily operations brief — with clear rules and a person who reviews the first few weeks of output closely.

Cheaper models make this affordable to run around the clock, and the platforms now provide the plumbing. The work left for a business is the part that was always its own: deciding what good looks like, where a human must sign, and how to know when something has gone wrong.

Key takeaways

  • Always-on agents are now mainstream products (OpenAI’s dots, Microsoft’s Autopilot), not research demos.
  • Each one ships with a permission model — allow, require approval, block — and that policy is the heart of the product.
  • Build your own automations the same way: autonomous for routine reversible steps, a human gate for money, access and customer contact, and a full audit trail.

Choose how to get started

Build with us

A system engineered around how you work.

Scoped in a proposal, designed as a prototype, delivered in milestones you can review.

Start a project

See the work

Systems already running in production.

Marketplaces, operations portals and mobile apps, with the case studies behind them.

View our work
WhatsApp