Long Horizon Workflows

# Long-Horizon Workflows

Part of [Agent Skills™](https://github.com/itallstartedwithaidea/agent-skills) by [googleadsagent.ai™](https://googleadsagent.ai)

## Description

Long-Horizon Workflows enable AI agents to execute multi-hour, multi-phase autonomous pipelines that far exceed the scope of a single conversation turn. Inspired by the DeerFlow architecture and production patterns from [googleadsagent.ai™](https://googleadsagent.ai), these workflows decompose complex objectives into staged execution plans with checkpoints, progress tracking, error recovery, and human-in-the-loop gates at critical decision points. A long-horizon workflow might analyze an entire Google Ads account (dozens of campaigns, thousands of keywords), generate a comprehensive optimization report, and prepare implementation-ready change sets — all autonomously over several hours.

The core challenge of long-horizon execution is reliability. A workflow that takes 3 hours but fails at hour 2.5 with no recovery is worse than useless — it wastes time and compute. Checkpoint management ensures that work completed before a failure is preserved and can be resumed. Progress tracking provides visibility into what the agent is doing and how far along it is. Human-in-the-loop gates allow a human to validate critical decisions (like budget changes) before the agent proceeds, preventing catastrophic errors in unattended operation.

DeerFlow's contribution to this pattern is the concept of hierarchical task decomposition — a planner agent breaks the objective into phases, each phase into tasks, and each task into atomic operations. Each level of the hierarchy has its own checkpoint, timeout, and error handling policy. This creates a robust execution model that can survive individual task failures without losing the broader workflow state.

## Use When

- The objective requires more computation than fits in a single agent conversation
- Multi-campaign or multi-account analysis needs to run unattended
- The workflow has natural phases with different tool and data requirements
- Critical decisions require human approval before proceeding
- Long-running workflows must survive interruptions and resume cleanly
- You need audit trails showing what the agent did, decided, and produced at each stage

## How It Works

```mermaid
graph TD
    A[Objective] --> B[Planner Agent]
    B --> C[Phase 1: Data Collection]
    B --> D[Phase 2: Analysis]
    B --> E[Phase 3: Recommendations]
    B --> F[Phase 4: Report Generation]
    
    C --> G[Checkpoint 1]
    G --> D
    D --> H[Checkpoint 2]
    H --> I{Human Gate}
    I -->|Approved| E
    I -->|Rejected| J[Revise Analysis]
    J --> D
    E --> K[Checkpoint 3]
    K --> F
    F --> L[Final Output]
    
    M[Progress Tracker] --> C
    M --> D
    M --> E
    M --> F
    
    N[Error Recovery] --> C
    N --> D
    N --> E
    N --> F
```

The planner agent receives a high-level objective and decomposes it into ordered phases. Each phase executes a distinct part of the workflow (data collection, analysis, recommendation generation, report writing). Checkpoints between phases persist the intermediate state so the workflow can resume after interruptions. Human gates pause execution at high-stakes decision points, waiting for explicit approval before proceeding. A progress tracker provides real-time visibility into the workflow's status. Error recovery at each phase can retry failed tasks, skip non-critical tasks, or escalate to the human operator.

## Implementation

**Workflow Definition:**

```typescript
interface WorkflowPhase {
  id: string;
  name: string;
  tasks: WorkflowTask[];
  checkpoint: boolean;
  humanGate: boolean;
  timeout_ms: number;
  onError: "retry" | "skip" | "abort" | "escalate";
}

interface WorkflowT