A3S Code

A3S Code is an AI coding agent framework built around an intent-detection-first paradigm. Each user input is automatically classified into one of 8 intent types (locate, understand, retrieve, explore, reason, validate, compare, track), which dynamically switches the agent's operating style (general / explore / plan / verification / code_review) and injects targeted context to avoid the precision loss of one-size-fits-all approaches.

The framework provides 15 built-in tools (file read/write/edit, search, bash, web access, git worktree, parallel tasks, etc.), supports a Lead-Worker-Reviewer multi-agent collaboration model, and includes built-in skills such as agentic-search, code-search, code-review, explain-code, and find-bugs. On the security front, the AHP (Agent Harness Protocol) with 19 harness points enables external governance, combined with allow/deny permission policies, circuit breaker (stops after 3 consecutive LLM failures), auto-compact (context compression before token limits), HITL confirmation, taint tracking, PII redaction, and MicroVM sandbox isolation for enterprise-grade deployments.

The core is implemented in Rust (~79.5%), with native Python and Node.js bindings via PyO3 and napi-rs. It uses HCL for configuration, supports 20 trait-based extension points for replacing any component (LLM client, tools, memory, hooks, etc.), and integrates the MCP protocol. Four memory types (Episodic / Semantic / Procedural / Working) support cross-session persistence, Lane Queue provides priority-based task scheduling (P0→P1→P2→P3), and the planning system supports task decomposition with wave execution.

Architecture:

Agent (facade, config-driven, workspace-independent)
  └── AgentSession (workspace-bound)
        └── AgentLoop (core execution engine)
              ├── IntentDetector → ContextPerception (AHP 2.3)
              ├── ToolExecutor (15 built-in tools)
              ├── SkillRegistry / PluginManager
              ├── Planning + SessionLaneQueue
              ├── HookEngine (12 lifecycle events)
              ├── AHP Executor (19 harness points)
              ├── Security (permissions, taint, HITL, PII)
              ├── Context (RAG providers)
              └── Memory (4 types)

Installation:

pip install a3s-code   # Python
npm install @a3s-lab/code  # Node.js

After configuring an LLM provider in agent.hcl, initialize via Agent.create("agent.hcl"), create a session bound to a workspace, and start sending prompts.

Unconfirmed: The entry repo A3S-Lab/Code is a fork of upstream AI45Lab/Code (~10 commits behind, at v1.9.3 while upstream is v1.10.0); fork sync strategy is unclear. The a3s.dev/docs/code domain referenced in the README is unreachable (DNS error); the actual working docs site is on GitHub Pages. PyPI/NPM package pages were not directly verified. The relationship between A3S-Lab and AI45Lab organizations is not publicly documented.