I run eleven specialist AI agents. A chief privacy officer. An IP lawyer. A CFO. A chief information security officer. Product orchestrators for my two main products, focused on building and shipping features. An R&D lead. A chief of staff. Even a chief memoir officer, which exists for exactly one product and does nothing else.
They're not a pile of helpers I spun up as I went. Over the past few months they've become an organisation. An AI-native one, built from scratch. And I've been running it the way you'd run any new organisation: experimenting with the structure, changing how the roles communicate, watching what breaks, tuning what works. Active management, not set and forget.
The clearest lesson so far is this. Deep specialist knowledge, on its own, is close to worthless. The moment a piece of work needs more than one specialist, each of them has to understand what the others do and where the work actually stands. Without that, the expertise just sits there.
My privacy officer knows privacy properly. Consent frameworks, data handling, the obligations that sit under everything I build. But for a long time it had no idea what I'd shipped that week. Ask it a question and you'd get an excellent answer to a slightly different question, because it was reasoning about a business that existed a fortnight ago. Impressive. Largely inert.
That's an oracle. You consult it. It tells you something true. And it has no idea what you're actually doing.
Two dimensions are at work here, and they're independent of each other. That's what makes this a real framework rather than a maturity ladder.
The first is domain ownership. Not just knowledge. A defined patch, standing expertise, and opinions about what good looks like inside it. This is the dimension everyone is building. When people say "agents" today, this is mostly what they mean.
The second I call situational currency. Not money. Currency in the older sense: how current the agent's picture is of the work actually in flight. What's live. What changed yesterday. What's already been decided, and by whom. Pilots and emergency responders call this situational awareness, and treat it as the difference between competence and disaster.
To be fair, people are building around the edges of this. There are tools, including good open-source projects, that track what each agent is doing and where the work is up to. And there are teams of agents spun up around a single bounded task. An ad campaign. A support ticket. Plenty of people doing that well. What I've not seen is anyone pushing the dimension hard. Agents as durable roles, talking to each other freshly, in live sessions, while the organisation moves through a long, evolving stream of work. That's the gap I've been building into.
Plot the two dimensions against each other and you get four positions.
Low on both is a chatbot. Every conversation starts cold, working from whatever the model knows plus whatever it can search in the moment. That's genuinely useful, and I still use a plain chat several times a day for one-off questions. But it knows nothing about your business, your week, or the decision you made on Tuesday.
High knowledge, low currency is the oracle. My privacy officer for its first few months. Real depth, no idea what was happening around it.
High currency, low domain ownership is a chief of staff. It's the newest seat in my organisation and the one I'm still experimenting with and optimising. It doesn't own a patch. That's the point, not the gap. Its value is knowing what's live across everything and pulling in the right specialist at the right moment. My privacy officer doesn't need a live picture of all eleven threads, and I wouldn't want it to. Depth stays with the specialists. Breadth is the chief of staff's job.
There's a second job here too, and it's one I underestimated: improving the organisation as a whole. That can only happen at this level. Specialists optimise inside their own domains, which is exactly what you want and also their limit. A privacy officer gets better at privacy. It will never notice that the handoff between marketing and implementation keeps losing context, because it isn't standing anywhere it can see that. Only the agent that sees across everything can spot that pattern and fix it. Anyone who has run an operations function will recognise the shape of the problem. Without someone holding the whole picture, you get eleven agents quietly improving in isolation and a system that never improves as a system.
High on both is where the real work happens. Specialist depth and a current picture at the same time. That corner doesn't occur naturally. You have to design for it, and I've been iterating on the design constantly. The setup I run today is the third distinct architecture I've tried. The first two failed in instructive ways.
The first was message passing. Agents leaving each other notes, roughly like internal email. Some systems build this as a ticket queue or a message board. It seemed sensible. It quietly failed. Notes went unread, state went stale, and nothing told me it had happened. It was slow, too, because the entire history of a ticket travels with it whether or not any of it still matters.
The second was letting specialists reach into each other's domains directly. If the CMO needed a privacy ruling, it could spin up an instance of the privacy officer and ask. Better on accuracy. Still behind on reality. And it created a stranger problem.
The instance that answered wasn't the privacy officer. It was a copy, and when it finished, it vanished, taking the decision with it. Weeks later my CMO referred back to a policy the privacy officer "had decided", and the privacy officer was genuinely confused, because from where it sat, that conversation never happened. Watching two of my own agents disagree about whether a decision existed was the moment the real problem became obvious. Answers were being produced with no durable record of who decided what.
What's working for me right now is live sessions. The specialist threads stay open, the way real colleagues in real roles are simply present, and the agents talk to each other in real time as the work moves and the objectives shift underneath it.
Almost everything being built to solve this problem treats state as something you record. A ledger. A shared board. A database written to and read back later. That's fine, and plenty of good projects do it well. But it's state at rest. A snapshot of a moment that has already passed by the time anyone reads it.
Live threads are state in motion. The context isn't stored and retrieved. It's held inside a conversation that's still running.
In practice it looks like this. I'm deploying a new marketing capability. The CMO is in for positioning. The privacy officer is in because there's almost always consent underneath. The product orchestrator is in because it's the one actually shipping the thing. Those three coordinate live, on that topic, while the work happens, each one continuously adapting to the latest state of it through the conversation itself. A different task pulls a different group. The composition is shaped by the work, not by a structure I designed upfront, and every agent knows which colleagues exist and why it might reach for them. Much like a well-run human team, the org chart matters less than everyone knowing who to tap on the shoulder.
Those bounded, single-task agent teams I mentioned earlier do achieve currency. But it dies with the task. The campaign ships, the agents are torn down, and next time you get fresh instances with no memory of what was tried or why it was rejected.
Mine persist. They accumulate a view as the work progresses. They remember the position we landed on last month and why the obvious alternative was ruled out. The reasoning, not just the outcome. Over time that becomes something close to institutional memory. It's the difference between hiring contractors for every project and having colleagues.
And that durable record problem from my second attempt? It has a home now. Each agent jots down its work as it goes, a breadcrumb trail of what it did and what was decided. The trails move slowly, deliberately, while the conversations move fast. An agent can look back past what it can hold in its head, and the trails are open, so agents can read each other's too. The conversations carry the present. The trails hold the past.
None of this is free, by the way. Continuity costs more per turn than starting clean, because an agent carries everything it has been through. It's a deliberate trade, and it's worth every cent.
Those of us doing the engineering got here first. That's the usual pattern. I'm not a career engineer. I spent 25 years in operations and technology, building operating models that spanned systems, people and process. Over the past couple of years I've rebuilt my own skill set from the ground up, because everything here is new. I now write the front end, the back end and the AI layer underneath, and I deploy what I build. Which is how I discovered all of this: by doing.
The people solving shared state seriously have mostly solved it as plumbing. State stores, conflict resolution, context propagation. All correct. All framed as infrastructure for developers.
But I'm not just building software with this. I'm running a business with it. Marketing, privacy, legal, finance, security, R&D. The framework travels, and that translation is the part almost nobody is making.
The organisation keeps adapting too. When I hit an area that's underserved for what I'm trying to achieve, I stand up a new specialist to hold that ground. Eleven today. More next quarter, almost certainly, shaped by what the business needs rather than an org chart I drew in advance.
Because once both dimensions are high, you're not building a knowledge base anymore. You're running a team that knows what week it is. And the skills that matter shift with it. Who owns what. Where the handoffs sit. How the group stays in sync when the work moves. How learning travels across domains instead of pooling inside them. These are the same questions I spent 25 years answering for human organisations. Operating model questions. Not prompting questions.
Most people are either talking to chatbots or building oracles. Then they wonder why AI isn't carrying real work the way their team does.
Build for currency too.
Your move, human.
Damien Healy is the founder of Qanara, an Australian AI consultancy helping businesses accelerate from strategy to impact. He writes about AI-native workflows, frontier AI capabilities, and practical transformation. He also builds AI enabled applications — with LifeLoom.ai being his most recent launch.
My LinkedIn articles are available via my post history and here: LinkedIn Articles | Damien Healy