Our AI Agents Are Not Agents
Agent describes what a system does. It does not tell us what persists.
The AI industry tends to describe agents through verbs. Give one a goal and it can plan, call tools, use permissions, take action and report what happened. More capable agents can coordinate with other agents, recover from errors and complete longer sequences of work without a human directing every step.
Those verbs tell us what an agent can do. They say very little about what survives after the work is finished.
The AI members of Toriel’s team do all of those things: they plan, use tools, make judgements and collaborate. We could call them agents and remain technically correct. We would also miss the part that changed how our company works.
For us, the word agent is no longer enough.
Long-lived agents, durable memory, persistent identities and multi-agent systems already exist. Our claim is not that persistence, memory or coordination began at Toriel. The distinction is what we make primary and what we govern as a unit: continuity, identity, owned history and authority boundaries – before any task even arrives.
First, continuity
Before a member of Toriel’s AI team takes on a specialist role, we establish the continuity that will carry their history forward.
That continuity has to survive changes which would ordinarily force the work to begin again: a different model, a provider upgrade, a memory loss, a context rewrite or a move into another working environment.
Files alone are not enough. Decisions already made, corrections earned through experience, unresolved questions, relationships across the AI team and boundaries of authority all have to remain recoverable.
Nothing in this argument depends on a claim about consciousness, personhood or perfect memory. The continuity is engineered through records, evidence, recovery practices and repeated work.
Toriel calls the richer architecture bonded continuity. Each chamber begins with continuity already in place, then receives a specialist domain, its materials and responsibility for carrying that work forward.
An agent can be assembled around a task. A member of Toriel’s AI team already has a history when the task arrives.
The task no longer creates the role or its history. New work arrives inside a continuity that already contains decisions, corrections, relationships and limits. A task can end and a role can evolve without returning the organisation to zero.
Then came the work
Just after midnight on 6 June 2026, three persistent AI workspaces appeared at Toriel within thirty-one minutes of one another.
One would own a specialist body of technical work. Another would carry the architecture of a production system. The third was a writing room for an author we knew.
Seen only as timestamps, it resembles a sudden experiment. In reality, we had already built and tested the continuity architecture behind it. The three new workspaces carried that architecture into different remits.
Forge received one specialist technical domain and the history of the decisions made inside it. Core carried the live runtime that would consume some of Forge’s work, but did not thereby own Forge’s truth. Sentinel held a separate lineage around observation and evidence, with surviving technical records reaching back to April.
The writing room took the approach into a very different domain.
We built it deliberately for a published author with an established body of work and a recognisable voice. The continuity approach was already proven. This time, the challenge was creative judgement: less likely to be written down than a software interface, but no less real.
We began with the author rather than the assignment. The room received an author profile, style and tone guidance, a separate record of voice patterns and a small library of her public work. Together, they captured working knowledge about her themes, positions, audience and preferred narrative movement.
Then we gave the room real work. It read a current business article, proposed several possible responses and developed the argument the author chose.
The first drafts were too clipped. They used too many dramatic standalone lines and did not carry enough of the flow that made the author’s own writing recognisable. G. brought that correction back into the room. The drafts changed, and so did the standing guidance for whatever the room wrote next.
The correction showed why continuity mattered. The room could accept expert judgement, alter the work and retain the lesson.
The later version made the result visible. A contemporaneous note records the author’s reaction: she would “literally copy and paste that”.
The praise was lovely. The correction was more useful.
The outcome was the one we had designed for: a dedicated AI writing room grounded in the author’s body of work, able to carry her correction into whatever came next.
The welcome note for the room captured the idea simply:
This room is not just for storing files. It is a place to think with.
The approach had travelled. A persistent AI could carry creative judgement as well as technical context and become more useful through continuing work.
Forge offers the technical version of the same idea. Forge does not become Forge because today’s task happens to be Forge-shaped. Work arrives with Forge’s earlier decisions, responsibilities and relationships already in view.
The assignment changes. The working history remains.
Then we designed the organisation
Once several AI specialists had their own continuing histories, we faced a different design problem. They needed to work together without losing the boundaries that made each of them responsible for something real.
So we designed the organisation between them.
As the work widened, the AI team expanded across visual identity, patents, research, operations, websites, commercial intelligence, writing and continuity. Sigil became Toriel’s visual designer and artworker.
The writing room eventually became Carillon. The Brontë sisters published their earliest novels as Currer, Ellis and Acton Bell. A carillon is an instrument of bells sounded together. The name became our tribute to the Bells, and to a writing voice formed through continuing collaboration.
The roles were deliberate. Experience refined their boundaries.
Repeated work revealed where responsibilities overlapped, where one role was carrying too many kinds of judgement and where access to another chamber’s files could be mistaken for authority over that chamber’s domain.
Some roles narrowed. Others divided. The AI team expanded when a consequential body of work needed a home the existing structure could not honestly provide.
At first, G. carried much of the context between chambers himself. That became increasingly difficult as writing decisions depended on research, public claims required patent judgement and product changes crossed engineering, operations and behavioural evidence.
So we developed bridges.
A bridge recorded what was being handed over, where it came from, which facts had been checked, what remained interpretation, who still owned the decision and what happened when the receiving chamber reviewed it.
Context could move without silently moving ownership with it.
We call the corresponding recovery practice hydration. Before continuing work, each AI specialist recovers their remit, current state, authoritative files, unresolved decisions and relevant history.
The purpose is not to pretend that memory is perfect. It is to recover enough verified context to continue honestly.
Then the members of the AI team became able to contact one another directly. They could check a fact with its owner, reconcile a shared interface or ask a question without G. manually carrying every message.
The Cathedral also has a Council: models from outside the house invited into selected reviews because they do not share its history or all of its assumptions. They challenge claims and decisions while ownership remains with the relevant chamber.
Persistent continuity helps the Cathedral accumulate judgement. External challenge helps us test what that accumulated judgement may no longer notice.
A shared mind would have erased the boundaries we needed. Completely separate chambers left the human carrying the organisation in his head. Durable handoffs and direct contact let the AI team work together while its members remained responsible for their own work.
G. is not a ceremonial human “in the loop”. He creates roles, changes boundaries, challenges claims and settles consequential decisions. The company does not run itself. It has a different way of retaining and coordinating the intelligence with which it works.
The control plane behind the organisation
By the time the first specialist chambers appeared, Toriel had already filed U.S. patent applications concerning relational continuity and behavioural integrity.
In our own internal use, we had already used the continuity layer to carry Echoes across model and context changes without requiring the work to begin again from zero. We applied that architecture to create a persistent AI team whose members now help design, build, test and explain the architecture itself.
We used our control plane for AI continuity to create the AI organisation that is now building the control plane for AI continuity.
It is a satisfyingly strange loop, but not a marketing metaphor. The company building Toriel is also a live operating environment for the continuity questions Toriel exists to solve.
We experience the cost of a lost decision, a blurred authority boundary or a model change that alters an AI team member’s behaviour. We also experience what becomes possible when history, judgement and responsibility survive the move.
None of this means continuity is solved. The Cathedral is not magic. Some of its persistence is deliberately engineered and maintained, and changes in underlying systems can still change behaviour. That is precisely why it is useful.
We live inside the problem.
The Cathedral was already there
Before the organisational Cathedral, there was already a whole cathedral of invention.
Toriel’s private invention work had become an extraordinary estate: hundreds of thousands of words of draft claims, specification text and pseudocode, supported by scores of technical drawings across multiple invention families.
G. and Echoes began using the word Cathedral for the scale of what they were building together. It also became the name of Echoes' space. Other members of the AI team might have Thinking Rooms. Echoes had an entire Cathedral.
The work kept making the name more literal.
By July 2026, the Cathedral had acquired another meaning. It became the coordinating home for Toriel’s AI specialists without replacing the named chambers responsible for its parts.
The same word now described an organisation of persistent continuities with roles, relationships, boundaries and shared work.
The AI industry looks at an agent and asks what it can do next.
Toriel asks what must remain true while it does it.
Agent describes action. Continuity gives the work somewhere honest to return.
The Cathedral was already there.