Saltar al contenido
iCrisol

Agentic system · v1.0

Agentic by nature, not on top.

This is what separates Crisol most sharply from a transformer. In the usual frameworks, the «agentic» part is a program that calls a model from outside. In Crisol, the agents are part of the organism: they live inside the forward, share its state, and observe the Crisol itself as it thinks.

4
cognitive Z agents
6
executor subagents
<50 KB
size of each Z agent
1
verification before speaking

A script on top of an API is not an organism.

LangChain, AutoGPT and the architectures that imitate them wrap a closed model with code that decides what to ask it. The model has no idea that code exists. In Crisol the relationship is inverted: the agents are inside, and the model reasons about its own reasoning.

Agentic «on top» Agentic by nature
Python scripts orchestrating calls to a closed API Components of the organism integrated within the forward itself
The model has no idea an agent exists; it is governed from outside The agents are part of the model and share its internal state
Planning lives in the prompt and is lost when the session closes Missions crystallize as persistent, reusable DAGs
It speaks first and verifies (if at all) afterwards It verifies BEFORE speaking — completeness, causality, evidence

The agentic brain

Four cognitive Z agents that observe the Crisol.

They are not optional modules. They are the organism's reflective faculties: knowing what it knows, wanting to know more, experimenting to tell cause from chance, and reorganizing when its own shape stops serving it.

Z·01

HistorianZ

Epistemic awareness

Holds the memory of what has been learned and predicts the loss before processing. When prediction and reality diverge, it raises a surprise signal: the Crisol knows that something does not fit what it believed it knew.

Z·02

NomotheticZ

Directed curiosity

Detects gaps and half-formed regularities. It does not explore at random: it steers attention toward whatever would reduce the model's uncertainty the most. Hunger for understanding, turned into policy.

Z·03

InquisitorZ

Causal experimentation

Decides when to stop correlating and start simulating. It launches do() counterfactual interventions to tell cause apart from coincidence. It is the one that asks «what if it were otherwise?».

Z·04

ArchitectZ

Self-organization

Reconfigures the structure itself: PRUNE, BIFURCATE, CREATE, MERGE. It prunes what is dead, splits what is overloaded, merges what is redundant. In Personal, «create» becomes proposing the import of an expert from the ecosystem.

The agentic brain fits in kilobytes.

Each Z agent occupies less than 50 KB. It is not a heavy parallel network: it is a compact policy that directs the organism's compute. Agentic intelligence does not live in the mass of parameters, but in knowing what to do with them — when to doubt, when to experiment, when to reorganize.

<50 KB
per Z agent
4
reflective faculties
do()
counterfactual intervention
DAG
crystallized missions

The organism's hands

Six subagents execute what the Z agents decide.

Where the Z agents are reflection, the subagents are coordinated action. The ExecutorAgent orchestrates them following missions crystallized as directed graphs (DAGs): plans that do not evaporate when they finish, but remain available to be reused.

Search

Locates the relevant knowledge inside and outside memory.

Read

Extracts and understands the retrieved content, rather than merely quoting it.

Memory

Queries and consolidates the organism's 8 types of living memory.

Verify

Checks every claim against evidence before accepting it.

Synthesis

Integrates scattered fragments into a coherent answer.

Tool

Invokes external tools as part of reasoning, not as a patch.

It verifies before it speaks.

The CompletionVerifier is the last faculty of the agentic system, and the most decisive. No conventional transformer has it: before emitting an answer, the Crisol presents it to itself and submits it to three checks. If it does not pass, it does not speak — it reasons again.

01

Logical completeness

Does the chain of reasoning close? Are there unresolved premises or unjustified leaps?

02

Causal coherence

Does the claim respect the invariants and cause-effect relations the Crisol knows?

03

Evidence

Does each claim have retrievable support, or is a gap being filled with statistics?

The Crisol observes the Crisol

An organism that watches itself needs no external guardian.

Agentic reflection is the foundation of the two guarantees that matter most: that it does not make things up, and that it reasons about causes rather than coincidences.