Universal Agent Fabric

The Ontology, Governance, and Composition layer for the Agent OS.

---

The Problem

You have agents. Dozens of them. Each one is a bespoke snowflake—prompts copy-pasted, tools re-wired, governance bolted on as an afterthought. When you need to update a safety rail, you touch 47 files. When you add a new capability, you forget three agents. Your "fleet" is a fragmented mess of YAML, JSON, and tribal knowledge.

This does not scale.

The Solution

Fabric is a library of possibilities. It applies the DRY principle to AI agents.

Instead of defining each agent from scratch, you compose them from reusable fragments:

Fragment	Purpose
Role	Who the agent is (persona, reasoning style, base graph)
Domain	What the agent knows (tools, vocabularies, context)
Policy	What the agent cannot do (governance, guardrails, budgets)

These fragments are compiled into a single, deployable manifest for the Universal Agent Kernel.

┌─────────────────────────────────────────────────────────────────┐
│                         FABRIC                                  │
│   "The Physics, Laws, and Biology of the Agent Universe"        │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                         KERNEL                                  │
│   "The Runtime. Executes what Fabric defines."                  │
└─────────────────────────────────────────────────────────────────┘

---

The Taxonomy

Fabric organizes agent definitions into four core abstractions:

◈ ARCHETYPES — manifests/roles/

Reusable agent personas. Not instances—classes.

An Archetype defines:

• The base graph template (planning loop, ReAct, debate)
• The system prompt skeleton
• The default capability set

# manifests/roles/researcher.yaml
name: "Senior Researcher"
base_template: "planning_loop"
system_prompt_template: |
  You are a Senior Researcher. Synthesize information deeply.
  Never provide surface-level answers. Always cite sources.
default_capabilities:
  - "web_search"
  - "read_file"

One Archetype. Fifty agents. Zero duplication.

---

◈ LAWS — policy/

Governance as Code. Applied at build time, not runtime.

Policies are injected into the manifest before deployment. Agents cannot bypass what they never receive.

# policy/rules/safety.yaml
- name: "no_trading"
  target_pattern: "execute_trade"
  action: "require_approval"
  conditions:
    risk_level: "high"

- name: "read_only_fs"
  target_pattern: "delete_file"
  action: "deny"

Security engineers own this directory. Prompt engineers never touch it.

---

◈ MATTER — ontology/

The capability definitions. Domain-specific tools and vocabularies.

A Domain bundles:

• Capabilities: Abstract tool definitions (protocol, config templates)
• System Prompt Mixins: Context injected into routers
• Vocabularies: Domain-specific terminology

# ontology/domains/finance.yaml
name: "Finance Domain"
description: "Tools and knowledge for financial analysis."

system_prompt_mixin: |
  You have access to real-time market data.
  All monetary values must be formatted as USD.
  Be aware of SEC regulations regarding forward-looking statements.

capabilities:
  - name: "get_stock_price"
    description: "Get current price for a ticker."
    protocol: "mcp"
    config_template:
      command: "mcp-finance"

  - name: "get_company_filings"
    description: "Retrieve 10-K/10-Q filings."
    protocol: "mcp"
    config_template:
      command: "mcp-sec"

Define "Web Search" once. Use it in 50 agents.

---

◈ GENESIS — composition/

The build system. Mixins and templates for stitching fragments together.

composition/
├── mixins/          # Injectable behaviors (observability, safety prompts)
└── templates/       # Workflow patterns (fan-out, checkpointing)

---

The Compilation Pipeline

flowchart LR
    subgraph Inputs
        R[Role YAML]
        D[Domain YAML]
        P[Policy YAML]
    end

    subgraph Fabric
        B[FabricBuilder]
    end

    subgraph Output
        M[Kernel Manifest]
    end

    subgraph Runtime
        K[universal_agent_arch]
    end

    R --> B
    D --> B
    P --> B
    B --> M
    M --> K

What happens during compilation:

1. Load — Parse YAML into Pydantic models (Role, Domain, GovernanceRule)
2. Assemble — Create FabricSpec (the universe definition for one agent)
3. Compile Policies — Convert governance rules to kernel PolicySpec
4. Compile Tools — Aggregate capabilities from domains; attach policy refs
5. Compile Routers — Merge role prompts with domain mixins; wire governance
6. Compile Graph — Instantiate the base template with router and tool nodes
7. Emit — Write kernel-compatible YAML/JSON

---

Usage

Installation

pip install universal-agent-fabric

For development (includes kernel for integration tests):

pip install universal-agent-fabric[dev]

The Build Command

fabric build \
  --role manifests/roles/researcher.yaml \
  --domain ontology/domains/finance.yaml \
  --policy policy/rules/safety.yaml \
  --name finance-researcher \
  --out dist/finance_agent.manifest.yaml

Or via module:

python -m universal_agent_fabric.cli \
  --role manifests/roles/researcher.yaml \
  --domain ontology/domains/finance.yaml \
  --policy policy/rules/safety.yaml \
  --out dist/finance_agent.manifest.yaml

CLI Reference

Flag	Required	Description
--role	Yes	Path to Role YAML
--domain	No	Path to Domain YAML (repeatable)
--policy	No	Path to Policy YAML (repeatable)
--name	No	Manifest name (default: generated-agent)
--out	No	Output path (default: manifest.yaml)

Deploying to the Kernel

# 1. Compile with Fabric
fabric build \
  --role manifests/roles/researcher.yaml \
  --domain ontology/domains/finance.yaml \
  --out my_agent.yaml

# 2. Install the kernel
pip install universal-agent-arch

# 3. Run the agent
cp my_agent.yaml manifest.yaml
uvicorn universal_agent.runtime.api:app --reload

---

Why Fabric?

Governance as Code

Policies are not suggestions. They are compiled into the manifest.

An agent built without policy/rules/safety.yaml literally cannot call delete_file. The capability is never wired. The guardrail is structural, not behavioral.

Composable Skills

ontology/capabilities/web_search.yaml  →  Used by 50 agents
ontology/capabilities/code_exec.yaml   →  Used by 12 agents
ontology/domains/finance.yaml          →  Used by 8 agents

Update the capability once. Recompile. Done.

Separation of Concerns

Directory	Owner	Responsibility
manifests/roles/	Prompt Engineers	Agent personas, reasoning styles
ontology/domains/	Domain Experts	Tools, vocabularies, context
policy/rules/	Security Engineers	Governance, guardrails, budgets
composition/	Platform Engineers	Build patterns, mixins

No one steps on anyone's toes.

Fleet Management

This is how you manage 1,000 agents without losing your mind:

1. Define 10 Archetypes
2. Define 15 Domains
3. Define 5 Policy Sets
4. Compose 1,000 unique agents from combinations
5. Update a guardrail in one file
6. Recompile the fleet

---

Directory Structure

universal_agent_fabric/
│
├── manifests/
│   ├── roles/              # Agent archetypes
│   │   └── researcher.yaml
│   ├── graphs/             # Graph topology templates
│   └── examples/           # Reference manifests
│
├── policy/
│   ├── rules/              # Atomic governance rules
│   │   └── safety.yaml
│   ├── templates/          # Parametric policies
│   └── validators/         # Policy validation logic
│
├── ontology/
│   ├── domains/            # Domain-specific bundles
│   │   └── finance.yaml
│   ├── capabilities/       # Abstract tool definitions
│   └── protocols/          # Protocol specifications
│
├── composition/
│   ├── templates/          # Composable workflow patterns
│   └── mixins/             # Reusable behavior fragments
│
├── universal_agent_fabric/ # Python package
│   ├── schemas.py          # Pydantic models (Role, Domain, etc.)
│   ├── builder.py          # Compilation logic
│   └── cli.py              # Command-line interface
│
├── tests/                  # Integration tests against kernel schema
├── pyproject.toml          # Package configuration
└── README.md

---

Schema Reference

FabricSpec

The root specification for compiling an agent.

class FabricSpec(BaseModel):
    name: str
    role: Role
    domains: List[Domain]
    governance: List[GovernanceRule]

Output Format

Compiled manifests conform to the kernel's AgentManifest schema:

Fabric Concept	Kernel Schema	Notes
Role.system_prompt_template	RouterSpec.system_message	Merged with domain mixins
Domain.capabilities	ToolSpec[]	Protocol: mcp, http, local, subprocess
GovernanceRule	PolicySpec.rules[]	Actions: allow, deny, require_approval
Role.base_template	GraphSpec.metadata.extra.template	Preserved for traceability

---

Development

# Clone
git clone https://github.com/mjdevaccount/universal_agent_fabric.git
cd universal_agent_fabric

# Create venv
python -m venv .venv
.\.venv\Scripts\Activate.ps1  # Windows
source .venv/bin/activate      # Unix

# Install with dev deps
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

---

License

MIT

---

Fabric defines. Kernel executes.
Separate the universe from its physics.