Skip to main content

Overview

1. Project Positioning​

A2E (AI Agent Engine) is an engineering execution engine for the AI era. It focuses on turning conversational intent into executable workflows while ensuring that the execution process remains interruptible, resumable, and traceable.

The current version is positioned as Experimental: this documentation focuses on the architecture, engineering implementation, and the methods and outcomes of AI-assisted development.

2. Core Ideas​

A2E currently centers on three core ideas:

  • AI-driven (primary Codex): use a primary Codex as the central collaboration entry point for requirement understanding, tool invocation, and task progression.
  • Controlled workflow state: place execution inside an explicit state machine and event stream so that runtime behavior is observable, interruptible, and resumable.
  • The application as MCP: expose application capabilities through MCP tools/resources so AI can interact with the system in a standardized way.

At the current experimental stage, execution capabilities are prioritized around a closed loop of AI nodes, while the review mechanism handles high-risk or high-uncertainty decisions to balance automated execution with human takeover.

3. Operating Model​

The current A2E operating model can be summarized as a controlled execution chain: workflow -> run -> node_run -> review -> resume. The system forms a minimal closed loop with two node types, agent.codex and human.review, while Web, Server, and MCP work together as the three entry points for orchestration, execution, and human intervention.

The operating model can be divided into three stages:

  • Orchestration stage: users create, edit, and save workflow definitions through Web or MCP. The Server persists workflow, node configuration, AI node metadata, and contract data, and prepares the runtime directory and compatibility-layer boundaries.
  • Execution stage: when workflow run is triggered, the Server creates run and node_run, and invokes the local codex exec inside AI nodes to complete the task. During execution, events, logs, artifacts, and structured outputs are continuously written for state tracking, result aggregation, and failure recovery.
  • Review stage: when a result requires human confirmation, or the flow enters a human.review node, the runtime state becomes paused and a review request is generated. After the review decision is made, the workflow can continue through resume.

The corresponding system responsibilities are:

  • Web: provides workflow editing, runtime monitoring, review handling, and the Codex interaction entry point.
  • Server: handles scheduling, state persistence, event recording, review coordination, and AI node runtime control.
  • MCP: exposes tools related to workflows, runs, reviews, and AI nodes to Codex or other MCP clients.

4. Current Scope and Boundaries​

The current minimal closed loop includes:

  • On the Web side: the home page, workflow list and editor, runtime monitoring, review list, settings, and other key page entries are already available.
  • On the Server side: APIs and service modules for workflows, runs, events, reviews, AI node execution and templates, and Codex PTY are already available.
  • On the MCP side: tools for Workflow / Run / Review / AI nodes are already available, aligned with Server semantics and capable of cross-end invocation.

5. Project Background and Current Trajectory​

The technical direction of A2E started from two stages of industry change:

  • In Q1 2025, the MCP paradigm gradually became one of the de facto standards for AI tool integration and capability orchestration.
  • Starting in Q2 2025, AI CLI capabilities and engineering adoption accelerated significantly, driving rapid evolution in AI-driven development architectures.

At present, trends represented by OpenClaw-style autonomous patterns and Skill-based capabilities continue to move forward. A single, static workflow paradigm has clear limits in complex scenarios. The future direction of A2E is to preserve the advantages of workflows being controllable and auditable while introducing more flexible AI-driven execution and execution boundaries that remain under control.