Skip to main content
This guide walks through building a reactive AI pipeline where changing one resource automatically propagates updates through the entire dependency chain.

​
What You’ll Build

A three-layer pipeline where:
  1. A secret stores an API key
  2. A model references the secret
  3. An agent depends on the model
When you change the secret, the model and agent automatically rebuild.

​
Prerequisites

  • pragma CLI installed and authenticated
  • Agno provider deployed (pragma providers list to verify)

​
Step 1: Define the Pipeline

Create a file called pipeline.yaml with all three resources: 
# Layer 1: Secret stores the API key
provider: pragma
resource: secret
name: anthropic-key
config:
  data:
    api_key: "sk-ant-your-key-here"
---
# Layer 2: Model references the secret's output
provider: agno
resource: anthropic-model
name: claude
config:
  model_id: claude-sonnet-4-5-20250929
  api_key:
    provider: pragma
    resource: secret
    name: anthropic-key
    field: outputs.api_key
---
# Layer 3: Agent depends on the model
provider: agno
resource: agent
name: assistant
config:
  model:
    provider: agno
    resource: anthropic-model
    name: claude
  instructions: "You are a helpful coding assistant."

​
Step 2: Apply Everything at Once

pragma resources apply pipeline.yaml
pragma-os resolves the dependency order automatically:
  1. The secret is processed first (no dependencies)
  2. Once the secret is ready, the model’s FieldReference resolves and the model is processed
  3. Once the model is ready, the agent’s Dependency resolves and the agent is processed
Watch the progress: 
pragma resources list

pragma/secret/anthropic-key [READY]
agno/anthropic-model/claude [PROCESSING]
agno/agent/assistant [PENDING]
After a few moments: 
pragma/secret/anthropic-key [READY]
agno/anthropic-model/claude [READY]
agno/agent/assistant [READY]

​
Step 3: Trigger a Change

Now swap the model. Update pipeline.yaml to use a different model ID: 
provider: agno
resource: anthropic-model
name: claude
config:
  model_id: claude-opus-4-6
  api_key:
    provider: pragma
    resource: secret
    name: anthropic-key
    field: outputs.api_key
Apply just the model: 
pragma resources apply pipeline.yaml
Since only the model’s config changed, pragma-os:
  1. Re-processes the model with the new model_id
  2. Detects that assistant depends on claude
  3. Automatically re-processes the agent with the updated model data
The agent rebuilds without you touching it.

​
Step 4: Inspect the Resolution

To see how pragma-os resolved the dependencies, inspect the resolved config: 
pragma resources describe agno/agent assistant -o json
The output shows:
  • config — Your original YAML with Dependency markers
  • resolved_config — The fully resolved version with actual model data embedded

​
FieldReference Patterns

​
Single Value References

Reference a specific output field: 
config:
  api_key:
    provider: pragma
    resource: secret
    name: my-secret
    field: outputs.api_key

​
Multiple References

A single resource can reference multiple dependencies: 
provider: agno
resource: agent
name: multi-tool-agent
config:
  model:
    provider: agno
    resource: anthropic-model
    name: claude
  tools:
    - provider: agno
      resource: mcp-tool
      name: code-tool
    - provider: agno
      resource: websearch-tool
      name: search
  instructions: "You have access to coding and search tools."

​
Debugging Tips

​
Check dependency states

If a resource is stuck in pending, its dependencies might not be ready: 
pragma resources describe agno/agent assistant
Look at the Dependencies section to see what it’s waiting on.

​
Compare config vs resolved_config

pragma resources describe agno/agent assistant -o json
  • If resolved_config is null, the resource is waiting for dependencies
  • If resolved_config has values, the resource has been submitted to the provider
  • Compare the two to verify references resolved correctly

​
Force re-processing

If a resource should have updated but didn’t, re-apply it: 
pragma resources apply pipeline.yaml
Idempotent apply means unchanged resources are skipped, and only modified resources are re-processed.

​
Next Steps

Reactive Dependencies

Deep dive into the dependency resolution system.

File Uploads

Add file resources to your pipelines.