Workflow Engine - OpenFang

The OpenFang workflow engine enables multi-step agent pipelines where each step routes work to a specific agent, and output from one step flows as input to the next. Compose complex behaviors from simple, single-purpose agents without writing code.

When to Use Workflows

Use workflows when you need to:

Sequential ProcessingChain agents in a pipeline (research → write → review)

Parallel ExecutionFan work out to multiple agents and collect results

Conditional LogicBranch execution based on previous step outputs

Iterative RefinementLoop until a quality gate is met

Quick Start

Create a workflow via the REST API:

curl -X POST http://localhost:4200/api/workflows \
  -H "Content-Type: application/json" \
  -d '{
    "name": "code-review-pipeline",
    "description": "Analyze code, review for issues, and produce summary",
    "steps": [
      {
        "name": "analyze",
        "agent_name": "code-reviewer",
        "prompt": "Analyze this code for bugs:\n\n{{input}}",
        "mode": "sequential",
        "output_var": "analysis"
      },
      {
        "name": "security-check",
        "agent_name": "security-auditor",
        "prompt": "Review for security issues:\n\n{{analysis}}",
        "mode": "sequential",
        "output_var": "security_review"
      },
      {
        "name": "summary",
        "agent_name": "writer",
        "prompt": "Write summary:\n\nAnalysis: {{analysis}}\n\nSecurity: {{security_review}}"
      }
    ]
  }'

Execute the workflow:

curl -X POST http://localhost:4200/api/workflows/{id}/run \
  -H "Content-Type: application/json" \
  -d '{"input": "Your code here..."}'

Step Configuration

Each step in a workflow has the following structure:

{
  "name": "step-name",
  "agent_name": "researcher",        // OR "agent_id": "uuid"
  "prompt": "Research: {{input}}",
  "mode": "sequential",               // sequential | fan_out | collect | conditional | loop
  "timeout_secs": 120,                // Default: 120
  "error_mode": "fail",               // fail | skip | retry
  "max_retries": 3,                   // Only for retry mode
  "output_var": "research_output",   // Optional: store output for later reference
  "condition": "ERROR",               // Only for conditional mode
  "max_iterations": 5,                // Only for loop mode
  "until": "APPROVED"                 // Only for loop mode
}

Agent Resolution

Specify exactly one of:

agent_name - Reference agent by name (first match)
agent_id - Reference agent by UUID

If the agent cannot be resolved at execution time, the workflow fails.

Execution Modes

Sequential (Default)

The step runs after the previous step completes. Previous output becomes {{input}}.

{ "mode": "sequential" }

Step 1

Receives initial workflow input

Step 2

Receives Step 1’s output as {{input}}

Step 3

Receives Step 2’s output as {{input}}

Fan-Out (Parallel)

Multiple steps run in parallel. All receive the same {{input}} from the last sequential step.

{ "mode": "fan_out" }

If any fan-out step fails or times out, the entire workflow fails immediately.

Example: Three agents brainstorm in parallel:

{
  "steps": [
    {
      "name": "creative-ideas",
      "agent_name": "writer",
      "prompt": "Brainstorm 5 creative ideas for: {{input}}",
      "mode": "fan_out"
    },
    {
      "name": "technical-ideas",
      "agent_name": "architect",
      "prompt": "Brainstorm 5 technical ideas for: {{input}}",
      "mode": "fan_out"
    },
    {
      "name": "business-ideas",
      "agent_name": "analyst",
      "prompt": "Brainstorm 5 business ideas for: {{input}}",
      "mode": "fan_out"
    }
  ]
}

Collect

Gathers all outputs from preceding fan-out steps. This is a data-only step (no agent execution).

{ "mode": "collect" }

Outputs are joined with \n\n---\n\n and set as {{input}} for subsequent steps. Typical pattern:

step 1: fan_out   →  runs in parallel
step 2: fan_out   →  runs in parallel  
step 3: collect   →  joins outputs
step 4: sequential →  receives joined output

Conditional

The step only executes if previous output contains the condition substring (case-insensitive).

{ "mode": "conditional", "condition": "ERROR" }

Example: Fact-check only if the article mentions claims:

{
  "name": "fact-check",
  "agent_name": "analyst",
  "prompt": "Fact-check this article:\n\n{{article}}",
  "mode": "conditional",
  "condition": "claim",
  "error_mode": "skip"
}

Loop

The step repeats up to max_iterations times. Terminates early if output contains the until substring.

{ "mode": "loop", "max_iterations": 5, "until": "APPROVED" }

Each iteration feeds its output back as {{input}} for the next iteration. Example: Iterative refinement until approval:

{
  "name": "review-and-refine",
  "agent_name": "code-reviewer",
  "prompt": "Review this draft. If quality is good, respond with APPROVED. Otherwise, provide feedback and a revised version:\n\n{{input}}",
  "mode": "loop",
  "max_iterations": 4,
  "until": "APPROVED"
}

Variable Substitution

Prompt templates support two kinds of variables:

`{{input}}` - Previous Step Output

Always available. Contains output from the immediately preceding step (or workflow’s initial input for the first step).

`{{variable_name}}` - Named Variables

When a step has "output_var": "my_var", its output is stored and can be referenced in later steps:

{
  "steps": [
    {
      "name": "research",
      "output_var": "research_output",
      "prompt": "Research: {{input}}"
    },
    {
      "name": "outline",
      "output_var": "outline_output",
      "prompt": "Outline based on: {{input}}"
    },
    {
      "name": "combine",
      "prompt": "Write article.\nResearch: {{research_output}}\nOutline: {{outline_output}}"
    }
  ]
}

Variables persist for the entire workflow run. Later steps can overwrite variables by using the same output_var name.

Error Handling

Fail (Default)

{ "error_mode": "fail" }

The workflow aborts immediately on error or timeout. The run state is set to Failed.

Skip

{ "error_mode": "skip" }

The step is silently skipped on error. A warning is logged, but the workflow continues. The {{input}} for the next step remains unchanged.

Retry

{ "error_mode": "retry", "max_retries": 3 }

The step is retried up to max_retries times after the initial attempt (4 total attempts). Each attempt gets the full timeout budget.

If all attempts fail, the workflow aborts with an error.

Complete Examples

Example 1: Code Review Pipeline

Sequential pipeline where code is analyzed, reviewed for security, and summarized.

{
  "name": "code-review-pipeline",
  "description": "Analyze code, review for issues, and produce summary",
  "steps": [
    {
      "name": "analyze",
      "agent_name": "code-reviewer",
      "prompt": "Analyze the following code for bugs, style issues, and security vulnerabilities:\n\n{{input}}",
      "mode": "sequential",
      "timeout_secs": 180,
      "error_mode": "fail",
      "output_var": "analysis"
    },
    {
      "name": "security-check",
      "agent_name": "security-auditor",
      "prompt": "Review this code analysis for security issues:\n\n{{analysis}}",
      "mode": "sequential",
      "timeout_secs": 120,
      "error_mode": "retry",
      "max_retries": 2,
      "output_var": "security_review"
    },
    {
      "name": "summary",
      "agent_name": "writer",
      "prompt": "Write a concise code review summary.\n\nCode Analysis:\n{{analysis}}\n\nSecurity Review:\n{{security_review}}",
      "mode": "sequential",
      "timeout_secs": 60
    }
  ]
}

Example 2: Multi-Agent Brainstorm

Three agents brainstorm in parallel, then a fourth synthesizes their ideas.

{
  "name": "brainstorm",
  "description": "Parallel brainstorm with 3 agents, then synthesize",
  "steps": [
    {
      "name": "creative-ideas",
      "agent_name": "writer",
      "prompt": "Brainstorm 5 creative ideas for: {{input}}",
      "mode": "fan_out",
      "timeout_secs": 60
    },
    {
      "name": "technical-ideas",
      "agent_name": "architect",
      "prompt": "Brainstorm 5 technically feasible ideas for: {{input}}",
      "mode": "fan_out",
      "timeout_secs": 60
    },
    {
      "name": "business-ideas",
      "agent_name": "analyst",
      "prompt": "Brainstorm 5 ideas with strong business potential for: {{input}}",
      "mode": "fan_out",
      "timeout_secs": 60
    },
    {
      "name": "gather",
      "mode": "collect"
    },
    {
      "name": "synthesize",
      "agent_name": "orchestrator",
      "prompt": "You received brainstorm results from three perspectives. Synthesize them into the top 5 actionable ideas, ranked by impact:\n\n{{input}}",
      "mode": "sequential",
      "timeout_secs": 120
    }
  ]
}

Agent refines a draft until it meets quality standards.

{
  "name": "iterative-refinement",
  "description": "Refine document until approved or max iterations",
  "steps": [
    {
      "name": "first-draft",
      "agent_name": "writer",
      "prompt": "Write a first draft about: {{input}}",
      "mode": "sequential",
      "timeout_secs": 120,
      "output_var": "draft"
    },
    {
      "name": "review-and-refine",
      "agent_name": "code-reviewer",
      "prompt": "Review this draft. If it meets quality standards, respond with APPROVED at the start. Otherwise, provide specific feedback and a revised version:\n\n{{input}}",
      "mode": "loop",
      "max_iterations": 4,
      "until": "APPROVED",
      "timeout_secs": 180,
      "error_mode": "retry",
      "max_retries": 1
    }
  ]
}

Event Triggers

The trigger engine provides event-driven automation. Triggers watch the kernel’s event bus and automatically send messages to agents when matching events arrive.

Creating a Trigger

curl -X POST http://localhost:4200/api/triggers \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "<agent-uuid>",
    "pattern": "lifecycle",
    "prompt_template": "A lifecycle event occurred: {{event}}",
    "max_fires": 0
  }'

Trigger Patterns

Pattern	Description
`"all"`	Matches every event (wildcard)
`"lifecycle"`	Matches any lifecycle event (spawned, started, terminated)
`{"agent_spawned": {"name_pattern": "coder"}}`	Matches when an agent with name containing pattern is spawned
`"agent_terminated"`	Matches when any agent terminates or crashes
`"system"`	Matches any system event (health checks, quota warnings)
`{"system_keyword": {"keyword": "quota"}}`	Matches system events containing keyword
`"memory_update"`	Matches any memory change event
`{"memory_key_pattern": {"key_pattern": "config"}}`	Matches memory updates where key contains pattern
`{"content_match": {"substring": "error"}}`	Matches events whose description contains substring

Prompt Templates

When a trigger fires, {{event}} is replaced with a human-readable event description:

"Agent 'coder' (id: <uuid>) was spawned"
"Agent <uuid> terminated: shutdown requested"
"Quota warning: agent <uuid>, tokens at 85.0%"
"Health check failed: agent <uuid>, unresponsive for 30s"

Use Cases

Health Monitoring

{
  "pattern": {"content_match": {"substring": "health check failed"}},
  "prompt_template": "ALERT: {{event}}. Investigate status of all agents."
}

Spawn Notifications

{
  "pattern": {"agent_spawned": {"name_pattern": "*"}},
  "prompt_template": "New agent created: {{event}}. Update fleet roster."
}

Quota Alerts

{
  "pattern": {"system_keyword": {"keyword": "quota"}},
  "prompt_template": "Quota event: {{event}}. Recommend action.",
  "max_fires": 1
}

Memory Changes

{
  "pattern": {"memory_key_pattern": {"key_pattern": "config"}},
  "prompt_template": "Config changed: {{event}}. Validate settings."
}

API Endpoints

Workflows

# Create workflow
POST /api/workflows

# List workflows
GET /api/workflows

# Execute workflow
POST /api/workflows/:id/run

# List workflow runs
GET /api/workflows/:id/runs

Triggers

# Create trigger
POST /api/triggers

# List triggers (optionally filter by agent)
GET /api/triggers?agent_id=<uuid>

# Enable/disable trigger
PUT /api/triggers/:id

# Delete trigger
DELETE /api/triggers/:id

CLI Commands

# List all workflows
openfang workflow list

# Create workflow from JSON file
openfang workflow create workflow.json

# Execute workflow
openfang workflow run <workflow-id> "input text"

# List all triggers
openfang trigger list

# Create trigger
openfang trigger create <agent-id> '"lifecycle"' --prompt "Event: {{event}}"

# Delete trigger
openfang trigger delete <trigger-id>

Limits and Quotas

Run Eviction: The workflow engine retains a maximum of 200 workflow runs. Oldest completed/failed runs are evicted when this limit is exceeded.

Step Timeout: Default 120 seconds per step (configurable)
Loop Iteration Cap: Default 5 iterations (configurable)
Hourly Token Quota: Enforced per-agent via max_llm_tokens_per_hour

Next Steps

Creating Agents

Build custom agents for your workflows

Event System

Learn about the event bus architecture

​When to Use Workflows

​Quick Start

​Step Configuration

​Agent Resolution

​Execution Modes

​Sequential (Default)

​Fan-Out (Parallel)

​Collect

​Conditional

​Loop

​Variable Substitution

​{{input}} - Previous Step Output

​{{variable_name}} - Named Variables

​Error Handling

​Fail (Default)

​Skip

​Retry

​Complete Examples

​Example 1: Code Review Pipeline

​Example 2: Multi-Agent Brainstorm

​Example 3: Iterative Refinement

​Event Triggers

​Creating a Trigger

​Trigger Patterns

​Prompt Templates

​Use Cases

Health Monitoring

Spawn Notifications

Quota Alerts

Memory Changes

​API Endpoints

​Workflows

​Triggers

​CLI Commands

​Limits and Quotas

​Next Steps

Creating Agents

Event System

When to Use Workflows

Quick Start

Step Configuration

Agent Resolution

Execution Modes

Sequential (Default)

Fan-Out (Parallel)

Collect

Conditional

Loop

Variable Substitution

`{{input}}` - Previous Step Output

`{{variable_name}}` - Named Variables

Error Handling

Fail (Default)

Skip

Retry

Complete Examples

Example 1: Code Review Pipeline

Example 2: Multi-Agent Brainstorm

Example 3: Iterative Refinement

Event Triggers

Creating a Trigger

Trigger Patterns

Prompt Templates

Use Cases

API Endpoints

Workflows

Triggers

CLI Commands

Limits and Quotas

Next Steps