# AutoGLM Agent System
A Claude Code multi-agent system modelled on [AutoGLM](https://autoglm.zhipuai.cn/) —
a foundation agent for autonomous GUI control through web browsers.
## Quick Start
```bash
# Start a GUI automation task
/gui-task Search for the cheapest flight from London to Tokyo next month
# Check if a specific action is safe
/safety-check Click the "Delete Account" button
```
## Architecture
```
User
└── task-planner ← coordinator: decomposes goal, orchestrates pipeline
├── action-validator ← safety: rates each step LOW / MEDIUM / HIGH
├── web-navigator ← executor: drives browser (navigate, click, type, extract)
├── human-checkpoint ← blocker: pauses for auth, payment, destructive actions
└── task-reporter ← output: abbreviated trace + final summary
```
## File Structure
```
.
├── CLAUDE.md # Project context for all agents
├── PRODUCT_OVERVIEW.md # AutoGLM product reference
├── .claude/
│ ├── settings.json # Hooks + permissions
│ ├── agents/
│ │ ├── task-planner.md # Master coordinator
│ │ ├── web-navigator.md # Browser execution engine
│ │ ├── action-validator.md # Risk assessor (read-only, Haiku)
│ │ ├── human-checkpoint.md # Human handoff handler
│ │ └── task-reporter.md # Final report generator
│ └── skills/
│ ├── gui-task/SKILL.md # /gui-task skill
│ └── safety-check/SKILL.md # /safety-check skill
└── scripts/
└── validate-action.sh # PreToolUse hook for web-navigator
```
## Agent Summary
| Agent | Model | Tools | Role |
|-------------------|--------|------------------------------------|-----------------------------|
| task-planner | Sonnet | Agent(...), Read, Write, Bash | Goal decomposition + coordination |
| web-navigator | Sonnet | Bash, WebFetch, Read, Write | Browser action execution |
| action-validator | Haiku | Read | Risk classification |
| human-checkpoint | Haiku | Read, Write | Human input handler |
| task-reporter | Haiku | Read, Write | Summary + trace generation |
## Human Takeover Triggers
The system automatically pauses and asks you for input when it encounters:
- Login / authentication / 2FA screens
- Payment or purchase confirmation
- Destructive actions (delete, unsubscribe, cancel)
- Any step the action-validator rates HIGH RISK
You can pause or stop the agent at any time.
## How It Mirrors AutoGLM
| AutoGLM feature | This system |
|------------------------------|------------------------------------------------|
| L4 autonomy (user as approver) | human-checkpoint pauses on blockers |
| Abbreviated action trace | task-reporter produces condensed trace |
| Sandboxed GUI execution | web-navigator + validate-action.sh hook |
| Human takeover for auth/payment | human-checkpoint agent + PreToolUse blocks |
| No cross-session memory | task-planner uses project-scoped memory |
## Logs
Agent activity is logged to `.claude/task-logs/agent-activity.log`.
Task reports are appended to `.claude/task-logs/latest.md`.