The Non-Anxious CTO
There is a meeting I have been in a few times now, in different rooms, with different people. Someone says, with feeling, "we have to do something about AI." Faces turn. Heads nod. A short silence. And then a list begins to form — tools to evaluate, vendors to meet, a training programme for the team, a policy on which models people are allowed to use for what, a working group, a strategy document, a board update. The list is well meaning. The list is also, almost without exception, the list of an anxious system trying to discharge anxiety by producing work.
I noticed something the last time I was in that room. I was not anxious. Not in a smug way. Not in a way that meant I was right and everyone else was wrong. Just genuinely, observably, not anxious. The faces around me were trying to find the next thing to do. I was trying to find the next thing to not do.
🕯️ Friedman, the underrated one
Edwin Friedman's A Failure of Nerve was published in 2007, from a manuscript he was still working on when he died. It is the most underrated leadership book of the last twenty years. It is not, strictly, a leadership book. Friedman was an American rabbi and family-systems therapist; he spent his career watching families produce the same dysfunctions over and over, and noticing that the dysfunction was always anxious before it was anything else.
His core observation, transplanted with very little adjustment to organisations, is that in anxious systems the leader's job is not to fix everyone else's anxiety. It is to be a differentiated, non-anxious presence. Anxious systems pull you in. They reward you for being reactive. They punish you for being calm, because your calm is mistaken for not-caring. The leader who survives intact does so by holding a steady self in the middle of the noise — not aloof, not detached, but visibly not running on the same fuel as the room.
I first read A Failure of Nerve four years ago and put it down. He has been enjoying a quiet renaissance since — in church and family-systems circles especially, where Mark Sayers' A Non-Anxious Presence has picked up the thread for a new audience. Friends I respect keep bringing him back up. I had not connected him to my day job until earlier this year, when I noticed that the management technique I was actually using on the AI question was not anything I had read in a CTO playbook. It was Friedman, slightly disguised.
📣 What the anxious system sounds like in 2026
Look at any of the major AI-for-CTOs pieces from the first half of 2026. The advice clusters in a few familiar shapes.
Manage your team's anxiety. People are worried they will be replaced. Talk to them. Don't get left behind. The half-life of competitive advantage is collapsing; move now.
Both are reasonable things to say. Both are also, in Friedman's frame, instructions to be more anxious — to take on more of the room's fear, internalise it, and discharge it as plans. The CTO who follows that playbook ends up as the most anxious node in the system. The team's anxiety, the board's anxiety, the regulator's anxiety, the consultant's anxiety — all of it routes through them. They become the place anxiety goes to be managed, which is exactly the role Friedman warns against.
I do not think the advice is wrong. I think it is incomplete in a specific way. It tells you what to do about AI without telling you who to be about AI. And the doing, in the absence of the being, produces a tired person with a longer list.
🧘 Non-anxious presence, applied
What is the alternative? In Friedman's framing, the non-anxious presence is not the absent presence. It is not pretending you do not care. It is the steady, visible refusal to be conscripted into the room's emotional fuel supply. You stay engaged. You stay curious. You ask questions that move the air. But you do not pick up other people's anxiety and carry it for them, because if you do, the system will keep producing more.
In a CTO context I think this looks like three things at once.
First, you stop treating AI as a threat to be managed and start treating it as architecture to be designed. Threats live in your stomach. Architecture lives on a whiteboard. Once a question has moved from "what will happen to us" to "what shape do we want to build," the anxious charge drops out of it. This is not denial. It is the Friedman move of reframing — a different angle of approach that changes what you are a presence to.
Second, you build slowly. AI moves fast; the architecture that endures has to move slowly. The temptation, when the rest of the room is moving fast, is to match speed. The Friedman move is to refuse — politely, visibly, with a reason. "We are not going to evaluate twelve agents this quarter. We are going to design the constitution under which any agent we ever adopt has to operate. Then we are going to evaluate three." The room will not enjoy this. The room is rarely the right judge of this.
Third, and this is the part I think most people miss, you make the architecture do the regulating rather than yourself. The whole point of building a system is so that you do not have to be present to keep it running. The Friedman leader is the one who has built a system that does not need them as the central node. Anxiety is what you get when everything depends on you reacting. Architecture is what you get when most things do not.
Non-anxious is not the same as slow. Some moments really do call for speed, and refusing to move in them is its own kind of failure. The point is not to reject urgency; it is to be able to tell genuine urgency apart from the emotional weather of the room. The architecture, when it is working, is what gives you the standing ground to make that distinction at all.
🏗️ Differentiation infrastructure
I am going to call the result differentiation infrastructure. The word "differentiation" is Friedman's. The infrastructure is mine, and I mean it literally — I have spent the last year building it, in small pieces, and I can be specific about what is in there.
A constitution for AI agents — a one-page document that says which combinations of authority no single agent will ever be allowed to combine. The Lethal Trifecta rule, as I have written about elsewhere: no agent should discover work, execute work, and verify its own work. The constitution does not say what the agents do; it says what they cannot. The shorter the better.
A shared, agent-readable memory — what I call cairn. Typed entries, edges, sensitivity tiers. Owned by us, in a Postgres database we control, exposed via an open protocol. Any agent can read it. Any agent can write to it. The agents share context the way a small team shares an office whiteboard, and the team and the agents share that whiteboard with each other.
Scoped roles — the agents are not "the AI." They are individuals, with names. Chief of Staff. Triage Engineer. Idea Scout. Financial Advisor. Security Analyst. Each has a defined remit, a defined set of tools, a defined audience. Each role is small enough that I could explain it to a new starter in a sentence.
A decision queue — when a role needs me, it does not interrupt. It writes a decision request to cairn with the proposed action and the rationale, and waits. I clear the queue when I clear the queue. The system does not generate panic; it generates a backlog I can work through.
None of that is technically difficult. The Postgres database, the protocol, the role files — those are a long weekend of careful work. The work that took longer, and the work that matters, was the upstream decision to architect this way rather than to chase capability.
The reason this is differentiation infrastructure rather than just engineering is what it does to me on a Wednesday afternoon. When the room is anxious about AI, I am not. I am not, because the architecture has already absorbed the question for me. The agents will or will not propose. I will or will not approve. The constitution will or will not let any of this go too far. The system has been designed to make my role a calm one — not because I am calm by nature, but because I have built around myself the thing Friedman is telling me to be.
🚧 The trap, named honestly
I want to be honest about where this gets hard, because it is too tidy otherwise.
The hard part is that other people read non-anxious presence as not-caring. Always. Without exception. If you sit in a meeting and someone says "we are losing the AI race" and you say, calmly, "I'm not sure that is the right framing," what they hear is "you do not care about the company." It does not matter how warmly you say it. The room is anxious; the room needs your anxiety as confirmation that the threat is real. When you do not give it, you become the problem.
Friedman calls this sabotage and he is unsurprisingly direct about it. Anxious systems sabotage their non-anxious members. They label them, isolate them, replace them. Sometimes the right response is to leave; more often the right response is to absorb the sabotage without absorbing the anxiety. Stay yourself. Stay curious. Keep building. Wait it out.
The other hard part is that you cannot fake this. You can copy the gestures — slow, steady, "let me think about that," low-volume meetings — and produce something that looks non-anxious from the outside but is just performative restraint. People can tell. The architecture either runs without you or it does not. The questions either route around you or they do not. The room will know within a quarter whether your calm is the real article.
I have not got this right. I notice anxiety in myself most days. I notice the reflex to discharge it as a plan, a meeting, a tool evaluation, a new document. I notice the gravitational pull of the room. The infrastructure helps. The infrastructure is not the whole answer. The whole answer is something closer to the long, slow practice of staying yourself, which Friedman would have called family-systems work and which I keep arriving at by a less rabbinical route.
🧭 What this amounts to
The Sum of All Tokens argued that software in 2026 costs the sum of token spend plus human insight, and that the insight is the scarce half. The Accidental Operating System argued that the organisational architecture AI agents needed was something like Holacracy — explicit roles, explicit decision rights, explicit constraint. This piece is what those two amount to when they meet a person.
The non-anxious CTO is not the unflustered one or the unaffected one. The non-anxious CTO is the one who has noticed that the room is producing anxiety as a substitute for thinking, and who has built infrastructure such that they no longer have to substitute. The architecture is the differentiation. The agents take the reactivity. The leader holds the long view.
If 2026 turns out the way I expect, the CTOs who survive intact will not be the most adoptive or the most aggressive. They will be the ones whose tech organisations are still recognisably themselves in three years. Differentiation is not a feature. It is a refusal — held steadily, in public, for long enough that the system around it learns how to work with a calm leader in the middle.
I would like that to be me. Some days it is. Some days I notice myself catching the room's pulse and have to put it back down. Friedman would say that is the work, and that the work is never finished. I think he is right about that as well.
James Webster is the founder of sheepCRM and director of Croftsware. This piece sits alongside The Accidental Operating System and The Sum of All Tokens.