agentlatch

Terminal-native agent resilience middleware for Python.

AgentLatch is a zero-dependency framework that makes AI agents resilient and observable. It solves two massive pain points in agent development:

1. Silent Tool Failures — When an LLM executes a tool that crashes, AgentLatch intercepts the Python exception, prevents a runtime crash, and feeds a structured JSON error back to the LLM so it can self-correct.

2. Blind Latency — It tracks millisecond execution time of LLM vs. tools and prints a color-coded ASCII flamegraph directly in the terminal using the rich library. No API keys, no dashboards, no cloud.

Quick Install

uv pip install -e ".[dev]"

Usage

from agentlatch import profile_agent, safe_tool

@safe_tool
def query_database(sql: str) -> str:
    """This tool is now protected — exceptions become JSON errors."""
    import sqlite3
    conn = sqlite3.connect("my.db")
    return str(conn.execute(sql).fetchall())

@safe_tool(timeout=5.0)
def call_api(url: str) -> str:
    """This tool has a 5-second timeout."""
    import requests
    return requests.get(url).text

@profile_agent
def run_agent():
    """The agent loop — traced and visualized automatically."""
    result = query_database("SELECT * FROM users")
    weather = call_api("https://api.weather.com/sf")
    return f"Got {result} and {weather}"

run_agent()

What Happens

1. Execution: Every @safe_tool call is timed and protected.
2. On Error: Instead of crashing, the tool returns a JSON error string:

   {
     "status": "error",
     "error_type": "ProgrammingError",
     "message": "column 'age' does not exist",
     "instruction": "The tool execution failed. Review your parameters and retry with corrected inputs."
   }
   

3. On Completion: A rich flamegraph is printed to the terminal:

   ┌─────────────────────────────────────────────────────────┐
   │  ⚡ AGENTLATCH EXECUTION PROFILE                        │
   │  Total: 1.23s │ Tools: 0.85s │ LLM Reasoning: 0.38s    │
   ├─────────────────────────────────────────────────────────┤
   │  ████████████████████████████████████████████ 1.23s      │
   │  ░░░░████████████░░░████░░░░░░░░░░░░░░░░░░░░            │
   │      query_db 0.5s  call_api 0.2s                       │
   │      ▲ ERROR                                             │
   │  Legend: █ LLM  █ Tool (OK)  █ Tool (ERROR)              │
   └─────────────────────────────────────────────────────────┘
   

Features

Running Examples

# Vanilla agent with a forced failure + self-correction
python examples/vanilla_agent.py

# LangGraph-style state machine
python examples/langgraph_agent.py

Running Tests

uv pip install -e ".[dev]"
pytest tests/ -v

Architecture

• contextvars — Thread-safe trace propagation without manual trace IDs
• concurrent.futures — Cross-platform timeouts (no signal.alarm)
• rich — Premium terminal rendering (the only external dependency)

License

MIT