Workflow Agents

Workflow Agents run multi-step pipelines with tool use. They chain multiple agents, branch on conditions, and call tools like document search, bucket operations, and web fetch. The policy gateway controls which tools each agent can use, and every step is logged for audit.

Multi-Step Pipelines

Workflow agents orchestrate sequences of steps:

1. Agent chaining — Output of one agent feeds into the next
2. Tool use — Agents call tools (search, fetch, bucket ops) and use results
3. Conditional logic — Branch based on step outcomes or user input
4. Parallel execution — Run independent steps concurrently when possible

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   Step 1    │────►│   Step 2    │────►│   Step 3    │────►│   Step 4    │
│  Document   │     │  Classify   │     │  Conditional│     │  Summarize  │
│  Search     │     │  (LLM)      │     │  Branch     │     │  (LLM)      │
└─────────────┘     └─────────────┘     └──────┬──────┘     └─────────────┘
       │                    │                  │
       │                    │                  ├──► Path A: Notify
       │                    │                  └──► Path B: Escalate
       ▼                    ▼
  Tool: search         Tool: none

Chaining Agents

Chain RAG, LLM, and custom agents in a single workflow:

{
  "name": "support-pipeline",
  "type": "workflow",
  "steps": [
    {
      "id": "search",
      "agent": "knowledge-base-rag",
      "input": "{{user_query}}"
    },
    {
      "id": "classify",
      "agent": "intent-classifier-llm",
      "input": "{{steps.search.output}}"
    },
    {
      "id": "respond",
      "agent": "response-generator-llm",
      "input": "Query: {{user_query}}\nContext: {{steps.search.output}}\nIntent: {{steps.classify.output}}"
    }
  ]
}

Step outputs are referenced via {{steps.<step_id>.output}}. The workflow engine resolves these before each step runs.

Conditional Logic

Branch based on step results:

{
  "steps": [
    {
      "id": "classify",
      "agent": "intent-classifier-llm"
    },
    {
      "id": "branch",
      "type": "conditional",
      "condition": "{{steps.classify.output}} == 'urgent'",
      "then": { "agent": "escalation-agent" },
      "else": { "agent": "standard-response-agent" }
    }
  ]
}

Supported condition operators: ==, !=, >, <, >=, <=, contains, startsWith, endsWith.

Policy Gateway

The policy gateway controls which tools each agent can use. Configure allow/deny rules per workspace, project, or agent:

{
  "policyGateway": {
    "rules": [
      {
        "tool": "document_search",
        "action": "allow",
        "scope": ["bucket:knowledge-base"]
      },
      {
        "tool": "bucket_list",
        "action": "allow",
        "scope": ["bucket:my-bucket"]
      },
      {
        "tool": "web_fetch",
        "action": "deny"
      }
    ]
  }
}

Tool Scopes

• allow — Agent can use the tool within the specified scope
• deny — Agent cannot use the tool
• scope — Restricts which buckets, URLs, or resources the tool can access

If a tool is not explicitly allowed, it is denied by default.

Audit Logging

Every workflow step produces an audit log entry:

Field	Description
stepId	Step identifier
agentId	Agent that ran
input	Resolved input to the step
output	Step output (or error)
toolCalls	Tools invoked and their arguments/results
timestamp	When the step completed
durationMs	Step execution time
userId	User who triggered the workflow
workspaceId	Workspace context

Use audit logs for:

• Debugging — Trace why a workflow produced a given output
• Compliance — Prove what data was accessed and when
• Analytics — Measure latency, tool usage, failure rates

Example Audit Entry

{
  "workflowId": "support-pipeline",
  "runId": "run_abc123",
  "stepId": "search",
  "agentId": "knowledge-base-rag",
  "input": "What is the refund policy?",
  "output": "Retrieved 5 chunks. Top result: ...",
  "toolCalls": [
    {
      "tool": "document_search",
      "args": {"query": "refund policy", "topK": 5},
      "result": {"chunks": 5}
    }
  ],
  "timestamp": "2026-03-01T12:34:56Z",
  "durationMs": 342,
  "userId": "user_xyz",
  "workspaceId": "ws_123"
}

Configuration Example

{
  "name": "research-assistant",
  "type": "workflow",
  "policyGateway": {
    "rules": [
      {"tool": "document_search", "action": "allow"},
      {"tool": "web_fetch", "action": "allow", "scope": ["https://docs.example.com/*"]},
      {"tool": "bucket_write", "action": "deny"}
    ]
  },
  "steps": [
    {
      "id": "search",
      "agent": "corpus-rag",
      "input": "{{user_query}}"
    },
    {
      "id": "synthesize",
      "agent": "synthesis-llm",
      "input": "Based on: {{steps.search.output}}\n\nAnswer: {{user_query}}"
    }
  ]
}

API: Run Workflow

curl -X POST "https://api.yourdomain.com/v1/agents/research-assistant/run" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "user_query": "Summarize our product documentation for new hires"
    }
  }'

Response

{
  "runId": "run_xyz789",
  "status": "completed",
  "output": {
    "steps": {
      "search": { "output": "..." },
      "synthesize": { "output": "..." }
    },
    "finalOutput": "Here is a summary of our product documentation..."
  },
  "auditLogId": "audit_abc123"
}

Next Steps

• Custom Agents — Build custom workflow logic and tools
• Agent Tools — Built-in tools and custom tool SDK
• RAG Agents — Use RAG as a step in workflows