| name | ais-agents |
| description | AI Agent framework — tool registry, multi-step reasoning, memory, rate control. Không infinite loop, tự động timeout. |
ais-agents — AI Agents
Core Principle
Agents must never infinite loop. Every loop must have a hard timeout and max steps.
Luật sắt
KHÔNG BAO GIỜ chạy agent loop KHÔNG có timeout + max steps
Mọi tool call phải có guard tràn token
Files
| File | Mục đích |
|---|
templates/tool-registry.py | Tool + ToolRegistry — type-safe tool registration, OpenAI schema |
templates/agent.py | Agent — ReAct loop with timeout (30s) and max steps (15), tool execution |
templates/agent-memory.py | AgentMemory — deque short-term + dict long-term, summarize() for context |
Usage
from templates.tool-registry import Tool, ToolRegistry
from templates.agent import Agent
registry = ToolRegistry()
registry.register(Tool(name="search", ...))
agent = Agent(registry)
result = await agent.run("find docs about X")
Security:
- Tool results are truncated to 2000 chars to prevent token explosion
- Agent timeout prevents runaway loops
- Validate all tool arguments before execution
Anti-Rationalization
| Excuse | Reality |
|---|
| "ReAct loop 15 steps là an toàn" | Tool call sai → loop tốn token, timeout |
| "Không cần memory summarization" | Context quá dài → LLM mất focus |
| "Tool result truncation 2000 chars thừa" | LLM context bị tràn bởi tool output dài |
Red Flags
- Agent không có timeout
- Tool args không validate
- Memory không giới hạn → token explosion
Verification
