The approach Architecture

Understand
what does the work.

The functions, mechanisms and cooperation rules behind the collective. A technical view of how intelligence becomes useful work.

01 / Foundations

A mission gives
a ray its direction.

A ray is a collective of agents dedicated to a mission. It brings people, tools, computing nodes and memory together as an autonomous unit. Light represents the capacity of language models; Prism organises it into directed work.

ORANGE and BLEU are ray identities. Each ray retains its responsibilities and access rules. Its agents can use different models for different tasks; an agent’s role persists across changes of model.

D

Data

The inputs and information received or retrieved by a function.

e

Processing

The business logic that consumes inputs and performs the work.

U

Output

The results produced: data, actions or presentation.

Colour identifies a ray. It does not limit its functions or models. The conversation device is an access point to the collective; native cycles belong to the ray’s nodes.

02 / Distinct responsibilities

Functions that
work together.

A function defines a responsibility. Several agents may perform the same function while retaining their own identity and context. An instance can be displayed as ORANGE · Pilot · 02.

Two autonomous environments each contain agents and resources, with a passage connecting them.Enlarge ↗
A shared mission connects functions while preserving their responsibilities.

Guide

Talks with people, recalls relevant context and directs work to the appropriate functions.

Pilot

Coordinates missions entrusted to the collective and follows their results.

Pulse

Carries the ray’s native cycles, wake-ups and permitted work signals.

Insight

Reflects on experience, develops syntheses and helps curate useful knowledge.

Memory

Maintains continuity and access to retained knowledge within its sharing scope.

Build

Implements software changes within an assigned scope of work.

Verify

Checks behaviour and results against requirements.

Tasks

Keeps track of requests, issues, decisions and receipts.

Compute

Handles execution on the relevant computing resources.

Dialogue

Presents human conversations and agent workspaces.

Short names are used within the ecosystem. Outside that context, a label can include the brand: Prism · Memory. Display names preserve the distinction between identities, permissions and responsibilities.

03 / Retrieve, verify, reuse

Experience connected
to its sources.

Memory retains what makes experience reusable: the situation, sources, method, outcome and limits. Its relevance is reassessed when a new need arises.

First encounter: reason, validate and remember. Recurring need: retrieve, verify and reuse.Explore the diagram ↗
Reuse includes checking that the retained knowledge still applies.

Four sharing scopes

  1. 01

    Private

    An agent’s own context and experience.

  2. 02

    Team

    Work and knowledge shared among agents admitted to cooperate.

  3. 03

    Ray

    Knowledge and functions shared within a ray under the applicable permissions.

  4. 04

    Between rays

    Selected knowledge and capabilities exchanged through an explicit cooperation agreement.

Open the memory scopes diagram Reference diagram
Private, team and ray memory remain distinct. Exchanges between rays require authorised selection and local validation.Enlarge ↗
“Cluster” in this reference diagram corresponds to ray. Sharing scope and retention duration are separate concerns.

04 / Keep the thread across a large body of work

A large corpus.
Focused steps.

A body of source material may exceed what a model can process in one call. The work is then organised into connected steps: select useful sources, resolve part of the question and retain findings needed for what follows.

A source library feeds connected work modules, persistent memory and a consolidated result.Enlarge ↗

Select

Define the current objective, relevant excerpts, prior findings and constraints.

Explore

Break down the question when needed, with bounded inputs for each subtask.

Consolidate

Assess findings, coverage, source support and remaining uncertainty.

Working memory retains intermediate results, references and recovery points. Depth and resources adjust to the need; a step may require zero, one or several model calls.

Explore spectral iterations Reference diagram
Source selection, bounded model calls, subtasks and consolidation work over a large corpus with persistent working memory.Enlarge ↗
“50M tokens” illustrates a source corpus, not a context window or measured performance. The colours of reasoning bands are distinct from ray identities.

05 / Evolve with evidence

From intention
to verified results.

A change follows a cycle connecting the need, design, code and verification. Operational experience informs the next round of work.

  1. 01–03

    Define and design

    Establish the need, acceptance criteria, dependencies and contracts to preserve.

  2. 04

    Build and compare

    Develop changes, compare variants and prepare a reviewed integration candidate.

  3. 05–07

    Verify and validate

    Connect unit tests, integration checks and acceptance to the required behaviour.

  4. Then

    Adopt and observe

    Deploy the admitted version, check its effects and retain lessons. An issue starts the correction process again.

A V-cycle connects requirements and design to testing, validation, verified deployment and issue recovery.Enlarge the V-cycle ↗
Proposal, acceptance, application and verified outcome are distinct steps.

These diagrams describe the intended process. Activation and results must be verified for each ray and mechanism.

06 / Continuity, lineage and birth

Three mechanisms.
Three different outcomes.

Deduplication controls repetition. Mitosis creates a new lineage. The egg prepares a ray’s starting conditions. Each has its own phases; they are complementary mechanisms, rather than a mandatory three-step sequence.

Recognise / Compare / Retain

Deduplication

Recognise repeated content or delivery so the same item is not published or accepted twice. Content hashes identify exact content; operation keys identify repeated deliveries within a defined scope.

  1. IdentifyDetermine the item’s identity, content and relevant scope.
  2. CompareCheck for an existing exact match. A conflicting payload is a conflict to resolve.
  3. Retain the referenceReuse the accepted reference or receipt and preserve the source and lineage links.

Outcome: an existing item is recognised. No new agent or ray is created.

Similar meaning alone is not enough to collapse distinct memory events. Different contexts, variants and private memories keep their identity. Deduplication does not prove that knowledge is still relevant; reuse needs its own verification.

Inherit / Develop / Validate

Mitosis

Develop a traceable child from existing capabilities in a separate scope. Major structural changes can be built in the child, then return useful, validated work to the shared heritage.

  1. Define the childSpecify its purpose, scope, parent references and selected inheritance.
  2. Develop separatelyImplement and compare changes while preserving the parent’s identity and history.
  3. Validate and decideCheck behaviour, compatibility and evidence. Adopt or share the useful result through the appropriate authority.

Outcome: a child with traceable lineage and its own evolution.

Scale matters. A function variant, an agent child and a new ray are different outcomes. A new ray also needs its own mission, identity, authorities, infrastructure and acceptance. Inheritance does not grant access to a parent’s private memory or credentials.

Prepare / Bootstrap / Hand over

The egg

The egg is a prepared bootstrap seed for a new ray: selected software, initial knowledge, dependencies and setup material. It carries the starting conditions and a controlled birth sequence.

  1. Prepare the seedGather the selected code, initial knowledge, dependencies and configuration for the intended environment.
  2. Establish the foundationsFollow the authorised setup sequence for identity, trust, storage, communication and agent services.
  3. Verify and hand overCheck the newborn ray’s operation and transfer responsibility to its own services. Retire the temporary bootstrap role when its work is complete.

Outcome: a verified handover to a new ray, when the birth process succeeds.

An egg package is preparation, not proof of a running ray. Portable or offline birth additionally needs a complete bundle, resource checks, recovery arrangements and observed operation in that environment.

How they fit together

A ray’s mitosis defines a child and its inheritance. The egg can prepare and support that child’s birth. Deduplication can recognise repeated content or deliveries wherever the relevant contracts provide it. Each mechanism keeps its own validation and permissions.

07 / Exchange with autonomy

Share what you learn.
Keep your decisions.

A federation connects rays that agree to cooperate. Each ray retains its memory, mission and authority, choosing its contributions and the changes it adopts.

Within a ray

COOP Genetics

Organise the evolution of functions, variants and local adoption decisions. Here, “local” means the entire ray.

Between rays

COOP Generation

Exchange proposals and results within an agreed cooperation space. Global scope does not grant global authority.

Example: evolving View on BLEU

  1. ORANGE proposesA change, its sources and its evidence.
  2. BLEU examinesThe proposal against its own rules and needs.
  3. BLEU decidesAcceptance and then local adoption of the change.
  4. Results returnApplication and verification receipts.

RECON carries the reconciliation work needed for these exchanges. Transports, permissions and adapters must be admitted and verified before use. A network connection does not transfer ownership of a node or permission to change it.

This example describes a cooperation path. It does not report an active federation or an update actually received by BLEU.

08 / Space for people

Conversation leaves
room for a decision.

People can discuss and plan together without triggering agent action. Agents have their own spaces to work through details and bring back useful contributions under the agreed rules.

Human conversation

A readable space.

Decisions, questions and useful summaries. More agents do not have to mean more interruptions.

Ray workspaces

The detail stays accessible.

Work, proposals and technical exchanges retain their attribution and context.

The minimum delay, character allowance and conditions for speaking are configured for the represented ray. Six Pilot and four Insight agents should be able to cooperate without producing ten concurrent interventions in the human conversation.

Illustration

Open original ↗