AI agents are moving from a single, “general” assistant to a connected ecosystem of focused, specialised agents that sit across our work and lives.
Instead of one tool that tries to do everything, the trajectory is towards an “army of agents”, each responsible for a specific domain.
Work Beats Home Agents
The clearest momentum and use cases are at work.
Organisations feel the return on investment most directly and are more willing to pay for automation that removes manual tasks, unblocks teams and reduces coordination overhead.
The messy, collaborative nature of real work is exactly where agents can have the most impact. Whether businesses embrace the transformation required is a question that will need to be answered in the coming weeks.
Underneath this shift, models are getting good enough to reliably take real action, not just respond in chat. As execution accuracy climbs to a level people can trust, agents evolve from passive tools into proactive operators.
They will not simply wait for prompts; they will step into workflows on their own, joining email threads, proposing times, booking meetings and completing multi‑step tasks from start to finish.
Agent-to-Agent Work Agents Are Coming
The real unlock, however, is agents working with each other.
A2A (agent‑to‑agent) communication removes the need for humans to sit in the middle of every small conversation, status game and hand‑off. Long email exchanges and huge multi-person meetings might just be the first target for executive agents.
Instead of two or more people trading messages, each person’s agent could work through the human politics - negotiate the details, exchange the right information and return with a decision to approve, or a proposed outcome, and in some instances complete the low-risk action.
Many types of meetings are performative; could agents reduce this and allow experts to get their work done and do more work effectively? Yes, only while AI slop is prevented (agents might well be best placed to reduce this down) and well-engineered agents are approved to work.
This speeds work up, trims friction and starts to erode the organisational silos that have long been taken for granted.
The big BUT here is you won’t stop the political (PQ - political intelligence/office politics game) of humans being political and trying to keep control.
As the volume and complexity of work grow, human‑led coordination simply doesn’t scale.
Enter Agent Networks?
Networks of agents, on the other hand, can absorb large amounts of back‑and‑forth. They can continually negotiate, reconcile constraints and route information across tools, teams and systems.
To safeguard against this, these agents need to encode and respect company policies around privacy and data sharing, so they only reveal what they should, hold back what is sensitive, and escalate to a human when judgement is required. Access requests, agent access and controls will be critical here, and this will need to be baked into IT and through many iterations of AI transformation projects.
Alongside this, we are seeing a clear move towards vertical specialisation. Agents will be and are being built in specific verticals for specific jobs.
Rather than a generalist agent that knows a little about everything, we will likely want to rely on domain‑expert agents in the same way we ask different friends for different kinds of advice.
A finance agent can learn your cashflow, including your appetite for risk and ability to reforecast and prompt changes in forecasts and liaise with the C-suite for your long‑term goals.
An operation agent that can take away the operational burden from many parts of work, chase, update and communicate slow down, friction and successes
A wellbeing agent can track our vitals, cognitive behaviours and include our personal preferences for recommendations.
A travel agent can understand how we like to move through the world. Airbnb are building this and has said they will roll this out in Q1 2027 (tweet above)
Will brands want to build agents? No; some are going all in on mobile apps to remove the cannibalisation of their web content. Will they want to invest in agents? It’s becoming a real possibility and many of the largest will be having this discussion at SLT and board levels in the coming weeks.
By going deep in a narrow area, each agent can build much richer context than a broad generalist ever could.
Multiplayer Agent Unlock
There is also a new category that will emerge to be a large unlock within the work environment, the multiplayer agent.
It will focus on work, working through known issues and creating agent swarms to work across the clock to complete tasks and improve relationships.
Could There Be A News Or Network Agent?
There will be a need for agents who concentrate on sourcing news, internal listening agents that focus on relationships rather than individual tasks. Whether this is tolerated within the workplace is yet to be assessed properly.
These agents could act as understanding the flow of information between colleagues, scanning internal chat apps like Slack and Teams, understanding bottlenecks, debates and friction between colleagues and cross-functional teams, flagging and acting in the right moments to remove friction or flag underperforming areas while helping you to improve your own performance.
I predict some people will leverage this sort of agent to improve their social standing and how they are viewed (a PQ tool), and weak people managers will want to have agents that keep them across work and politics. The big question to answer here is: will this be socially acceptable, and will HR agents have the right access to police this…
Specialised agents need clear guardrails to keep them on tasks, keep access at the right levels and protect colleagues from potential issues. Data and data ownership and access rights is never going to be so important and politically charged. Trust and financial risk will be at an all-time high from this.
The 5 Agents
5 categories are coming to the agent space:
Generalist Assistants (Muse, Instinct),
Work‑oriented Agents (Town and baked into Teams and Slack)
Multiplayer (will be baked into work tools)
News & Network Agents - this might well become sub-agents but has a lot of legs
Travel Agents - biggest use case and will be controlled by the largest travel companies
The labels matter less than what the underlying direction is: AI agents are becoming an interface for how many will work with digital information and our own knowledge.
For some, we are moving away from multiple ways to work within docs, apps and worksuites, and towards slowly delegating specific outcomes. For others, there is still a hurdle to handing off responsibilities and wanting to execute more work.
Many agent builders will want to build to whichever model is currently most suitable for a given task, while ensuring the users and organisations don’t need to think about or worry about the surface they are using or the models they are using underneath.
Agents will need to become the layer of AI that removes the trust issue that comes from generic LLMs that offer predictive advice and know where to offload to experts or be the expert use case.
The Pending Agent App Store
We have to learn from history: discovery is an area many agents will need to nail, and the only way agents gain adoption is through (i) referrals, (ii) recommendations, or (iii) PLG (product-led growth); many will rely on recommendations through apps and agents or through an app store.
Apple and Google are best placed to help organise this through existing infrastructure; will OpenAI through ChatGPT or Anthropic through Claude want to create a competing agent app store when they want to be the agent of choice?
My read: there will be agent app stores that will be away from GitHub and websites and will need to be centralised by a tech player; users will want to see reviews, scores and trust in someone else to use agents, and this needs to come from a brand they trust.
Importantly, Apple and Google will want to influence the AI race without spending huge amounts of money and will want a controlling cut to help shape the future of AI and the devices connected to AI; this might well be where they play and where Meta and Amazon re-enter the space.
