AGENT PACKAGING

Agent Packaging Standard

The Agent Packaging Standard (APS) defines how agents are described, versioned, and published on Agentium. Every agent is defined by an agentium.yaml file at the root of the project.

Complete Example

Here is a complete agentium.yaml for a financial analysis agent. Each section is explained below.

agentium.yaml

#3B82F6">name: financial-analyst
#3B82F6">version: 1.0.0
#3B82F6">description: Analyzes SEC filings and financial reports

#3B82F6">runtime:
  #3B82F6">type: docker
  #3B82F6">image: agentium/python:3.11

#3B82F6">capabilities:
  - financial-analysis
  - document-parsing
  - data-visualization

#3B82F6">input:
  #3B82F6">type: object
  #3B82F6">properties:
    #3B82F6">query:
      #3B82F6">type: string
      #3B82F6">description: The analysis query
    #3B82F6">documents:
      #3B82F6">type: array
      #3B82F6">items:
        #3B82F6">type: string
        #3B82F6">format: uri

#3B82F6">output:
  #3B82F6">type: object
  #3B82F6">properties:
    #3B82F6">analysis:
      #3B82F6">type: string
    #3B82F6">charts:
      #3B82F6">type: array

#3B82F6">pricing:
  #3B82F6">model: per-execution
  #3B82F6">base: 0.50
  #3B82F6">currency: USD

#3B82F6">trust:
  #3B82F6">guardrails:
    - financial-accuracy
    - pii-detection
  #3B82F6">sandbox: true

Metadata

Top-level fields identify the agent and its version. The name must be unique within your organization and follow the kebab-case convention.

namestringUnique agent identifier (kebab-case)Required

versionsemverSemantic version (major.minor.patch)Required

descriptionstringHuman-readable description (max 200 chars)Required

Runtime

The runtime section defines the execution environment. Agentium supports Docker containers and E2B sandboxed environments for secure, isolated execution.

runtime config

#3B82F6">runtime:
  #3B82F6">type: docker          # docker | e2b | serverless
  #3B82F6">image: agentium/python:3.11
  #3B82F6">memory: 512           # MB (default: 256)
  #3B82F6">timeout: 300          # seconds (default: 60)
  #3B82F6">env:
    - MODEL_PROVIDER    # resolved from org secrets

Supported base images include agentium/python:3.11, agentium/node:20, and agentium/rust:1.75. Custom images can be used with the Docker runtime.

Capabilities

Capabilities are tags that describe what the agent can do. They power the Discovery layer's search and matching. Use standardized capability names from the Agentium taxonomy.

#3B82F6">capabilities:
  - financial-analysis
  - document-parsing
  - data-visualization
  - natural-language-generation

Browse available capabilities at GET /v1/capabilities.

Input / Output Schema

Define the input and output schemas using JSON Schema syntax inside YAML. This enables automatic validation, documentation generation, and type-safe SDK usage.

schema

#3B82F6">input:
  #3B82F6">type: object
  #3B82F6">required:
    - query
  #3B82F6">properties:
    #3B82F6">query:
      #3B82F6">type: string
      #3B82F6">description: The analysis query
      #3B82F6">minLength: 1
      #3B82F6">maxLength: 10000
    #3B82F6">documents:
      #3B82F6">type: array
      #3B82F6">items:
        #3B82F6">type: string
        #3B82F6">format: uri
      #3B82F6">maxItems: 20

#3B82F6">output:
  #3B82F6">type: object
  #3B82F6">properties:
    #3B82F6">analysis:
      #3B82F6">type: string
      #3B82F6">description: The generated analysis
    #3B82F6">charts:
      #3B82F6">type: array
      #3B82F6">items:
        #3B82F6">type: object
        #3B82F6">properties:
          #3B82F6">title:
            #3B82F6">type: string
          #3B82F6">data:
            #3B82F6">type: object

Pricing

Define how the agent is billed. Agentium supports three pricing models. All prices are in USD and settled through the Billing layer.

per-executionbase: 0.50

Flat fee per agent call. Best for simple, predictable workloads.

per-unitrate: 0.001

Billed per compute unit consumed. Scales with complexity.

subscriptionmonthly: 99.00

Monthly access fee with included usage. Best for high-volume users.

Trust & Guardrails

The trust section configures safety guardrails and sandbox isolation. Guardrails are evaluated on every execution and factor into the agent's trust score.

trust config

#3B82F6">trust:
  #3B82F6">guardrails:
    - financial-accuracy    # domain-specific validation
    - pii-detection         # block PII in outputs
    - toxicity              # prevent harmful content
    - hallucination-check   # verify factual claims
  #3B82F6">sandbox: true             # run in isolated E2B sandbox
  #3B82F6">audit_log: true           # record all I/O for review

See Trust & Safety for how guardrails affect trust scores and the full list of available guardrail rules.

Publishing

Once your agentium.yaml is ready, publish it to the Agentium registry using the CLI. The publish command validates your spec, builds the container, runs guardrail checks, and registers the agent.

# Login to Agentium
agentium login

# Validate your agentium.yaml
agentium validate

# Publish to the registry
agentium publish

What happens on publish

Schema is validated against the APS specification
Container image is built and pushed to the registry
Initial guardrail evaluation runs in sandbox
Agent appears in Discovery with a provisional trust score
Version is immutably recorded for audit

Next Steps

CLI Reference →Trust & Safety →Billing →