Forge Agent: An Auditable, Model-Switchable AI Workflow Automation Platform

Project Core Information

• Original Author/Maintainer: jeremiaahhh
• Source Platform: GitHub
• Original Link: https://github.com/jeremiaahhh/ai-workflow-automation-agent
• Release Time: 2026-05-28

Core Views

This project is an AI workflow automation agent system built on FastAPI and Next.js, with the core design concept of "controllable autonomy". The system uses a plan→approval→execution three-stage workflow model. While ensuring AI automation capabilities, it guarantees operational controllability through manual approval nodes; it provides complete audit trails and Markdown report functions, and supports simulation mode operation without API keys to lower the trial threshold.

With the enhancement of large language model capabilities, AI agents have moved from concept to application, but enterprise-level deployment faces core contradictions: How to ensure the controllability, security, and auditability of operations when AI executes tasks autonomously?

• Fully autonomous AI agents are efficient but risky; wrong decisions may lead to serious consequences;
• Fully manual review loses the value of automation;
• The industry urgently needs a compromise: AI is responsible for planning and execution, key operations require manual confirmation, and the entire process is traceable.

Three-Stage Workflow Model

1. Planning Phase: AI analyzes task objectives, decomposes them into subtask sequences, identifies tools/resources, evaluates dependencies, predicts risks, and generates an execution plan (sandbox run with no actual operations);
2. Approval Phase: The execution plan is submitted for manual review, showing the execution path, tool parameters, resource consumption, and risk prompts. Operators can approve/modify/reject;
3. Execution Phase: Execute subtasks according to the plan, monitor status in real time, capture input/output, record time/resources, and trigger rollback/pause in case of exceptions.

Closed Tool Registry

The system only exposes predefined, security-reviewed tool sets (such as restricted file operations, read-only database queries, whitelisted HTTP requests, sandbox code execution) to reduce the risk of malicious operations.

Complete Event Log

Records all key events in the workflow lifecycle: task creation, plan generation, approval records, execution step status, exception handling, etc. Logs have timestamps and digital signatures to ensure they are tamper-proof.

Markdown Report Generation

Each completed workflow automatically generates a structured report, including execution summary, step breakdown, performance analysis, exception information, and original log links, which is easy to read and integrate (e.g., Git, document systems).

Backend: FastAPI

• Advantages: Type safety (Python type hints + Pydantic), high performance (Starlette + uvloop async), automatic OpenAPI documentation, dependency injection;
• Architecture: API layer (HTTP handling), service layer (business logic), agent layer (AI interaction), storage layer (data persistence).

Frontend: Next.js

• Advantages: SSR optimizes first-screen loading, ISR balances dynamism and performance, React component model, built-in API routes;
• UI Features: Plan visualization (flowchart), real-time status push (WebSocket), approval operation interface, Markdown report rendering/export.

Simulation Mode

When no API key is configured, the system uses predefined rules to simulate AI responses, provides sample tasks, marks "simulation mode", and supports configuring delay and randomness, making it easy to experience system functions without API permissions.

Enterprise Automation Workflows

• Data report generation: AI plans queries, manual confirmation for sensitive data access;
• Content publishing: AI drafts and formats, manual final review;
• Customer service tickets: AI classifies and provides initial responses, manual handling of complex cases.

DevOps

• Alert diagnosis: AI analyzes logs to plan troubleshooting steps, executes fixes after manual approval;
• Deployment process: AI generates plans, auto-executes after manual confirmation;
• Resource optimization: AI analyzes usage to provide suggestions, implements after manual decision.

Research and Experimentation

• AI agent security sandbox;
• Repeatable experiment environment (audit logs);
• Reference implementation of human-AI collaboration interface.

Environment Requirements

• Python 3.10+
• Node.js 18+
• Optional: Redis (state persistence/WebSocket), PostgreSQL (recommended for production)

Configuration Points

• TOOL_REGISTRY: Enabled tool whitelist;
• APPROVAL_MODE: Approval mode (always/on_error/never);
• MAX_RETRIES: Number of retries for step failures;
• AUDIT_RETENTION_DAYS: Audit log retention days.

Security Recommendations

• Enable approval mode in production;
• Regularly review the tool registry and remove unnecessary permissions;
• Off-site backup of audit logs;
• Follow the least privilege principle for API key configuration.

Summary

This project demonstrates a practical human-AI collaboration model, finding a balance between automation and controllability, suitable for enterprise-level AI applications (technology serves business, not a risk source).

Future Directions

• Intelligent approval prediction (auto-approve low-risk tasks based on historical data);
• Multi-agent collaboration and task allocation;
• Integrate more enterprise systems (ERP, CRM, BI tools);
• Enhance natural language interaction to lower the usage threshold.

For teams exploring AI agent implementation, this project provides a well-thought-out reference implementation worth studying and learning from.