Agent Specs — Chiralkit

Map · the whole chain

The chain

The Map composes complete agents into a workflow: handoffs, dependencies and mirrored builder/reviewer relationships. Each agent has its own internal chain, shown in the following rings.

Agent workflow

A task travels to a declared agent or agent group. Builder and reviewer can form a reflected pair here; each participant still has its own identity, memory, business domain, inference, technical capabilities, execution and consolidation. Ring positions on this Map follow current registry metadata; they do not represent an agent’s internal execution steps.

Ring 0 · atoms

Tokens

Ring 0: the agent’s foundation — purpose, function, identity, method, voice and persistent memory.

Identity, method and memory

Purpose, method, function, voice, identity and memory form the agent’s foundation. Follow the assembly or select a question to inspect one piece. Use the embedded controls to step through the scene and zoom in on its details.

Ring 1a · what the agent knows

Business domain knowledge

Ring 1a: one primary business domain gives the agent its vocabulary, relationships, rules, evidence and limits. Finance, livestock farming, automotive and product are examples of different domains.

What does this agent know?

Each agent has one primary business domain. The labelled membranes in the figure show alternative examples; select one below to inspect its layer. Concepts, associations, patterns and references remain connected perspectives on that knowledge.

Example business domains

Select an example to highlight its membrane in the figure. One primary domain per agent; the examples do not load every domain into one task.

Concepts

Concepts give business terms meaning: an invoice and its due date in finance, an animal and its identification in livestock farming, or a vehicle and its service record in automotive.

1a supplies business knowledge; 1b uses the relevant parts to interpret the task. Frontend, backend and DevOps are technical capabilities in ring 2. These groups illustrate connected knowledge, not anatomical brain regions or mandatory inference calls.

Ring 1b · what the task means

Contextual inference

Ring 1b: interpret a task, prompt or thought using identity, relevant memory and the business domain. Make requirements, constraints, uncertainties and acceptance criteria explicit before selecting technical capabilities.

Ring 2 · how the agent can act

Technical capabilities

Ring 2: select technical knowledge, skills, procedures, tools and checks for the interpreted task. Frontend, backend and DevOps are capabilities that can serve many business domains.

Frontend

Interaction, accessibility, interface architecture and browser testing. For a finance agent, these capabilities can turn business requirements into a clear invoice screen.

Domain answers what the task means in the business. Capabilities supply how to carry it out technically. Execution comes next; composition with another agent belongs in the Map.

Ring 3 · produce and verify

Execution

Ring 3: carry out a concrete operation using the contextual brief and selected capabilities. A technical spec, a code change or an evidence-based review is work performed here.

Define the concrete operation

Write a technical spec or bounded plan when the task needs one. Keep requirements, business rules, constraints and acceptance criteria attached to the work.

A completed operation returns an artifact and evidence, including unresolved failures. The architecture aims to reduce total tokens and cost per successful task; the local-model comparison has not yet been measured.

Ring 5 · retain verified learning

Consolidation

Ring 5: evaluate what execution taught the agent and decide what should update its persistent memory. Learning can concern business knowledge or technical procedures.

Evaluate the evidence

Separate verified results from guesses, failed attempts and observations that need more checking. A successful output alone does not justify every conclusion made along the way.

Consolidation updates external knowledge and procedures, not model weights. This visualization explains the intended architecture; a memory backend and consolidation policy still need implementation and validation.