agentdna-sdk 0.2.0

Sentry for AI Agents — one-line observability with local SQLite persistence

🧬 AgentDNA — Sentry for AI Agents

One-line observability for any Python agent. No framework required. No API key. No network calls.

pip install agentdna

Quick Start

from agentdna import observe, get_stats

@observe
def my_agent(prompt):
    # your existing agent code
    return llm.call(prompt)

my_agent("hello world")

# View stats (persists across restarts)
print(get_stats())

That's it. One decorator. Full observability.

What You Get

📊 Stats: my_agent
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Health:           ✅ Healthy
  Total calls:      42
  Success rate:     97.6%
  Failed calls:     1
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Avg latency:      1250.5 ms
  P50 latency:      980.0 ms
  P95 latency:      2100.0 ms
  P99 latency:      3500.0 ms
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Errors:
    Timeout: 1
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Features

• 📊 Call tracking — total calls, success/failure rates
• ⏱️ Latency — avg, p50, p95, p99 percentiles
• ❌ Error tracking — error types and frequency
• 💾 SQLite persistence — data survives restarts
• 🔒 100% local — no network calls, no API key
• 🐍 Sync & async — works with both
• 🖥️ CLI included — agentdna stats

CLI

agentdna stats                    # overview of all observed functions
agentdna stats my_agent           # detailed view
agentdna stats --export json      # export as JSON
agentdna stats --export csv       # export as CSV
agentdna stats --reset            # clear all data

Usage

Basic

from agentdna import observe

@observe
def my_agent(prompt):
    return llm.call(prompt)

With Options

@observe(name="transcriber", tags={"version": "2.0", "model": "whisper"})
def transcribe(audio_path):
    return whisper.transcribe(audio_path)

Async Support

@observe
async def my_async_agent(prompt):
    result = await llm.acall(prompt)
    return result

Get Stats Programmatically

from agentdna import get_stats

stats = get_stats("my_agent")
print(f"Success rate: {stats['success_rate']:.1%}")
print(f"P95 latency: {stats['p95_latency_ms']:.0f}ms")

# All functions
all_stats = get_stats()
for name, s in all_stats.items():
    print(f"{name}: {s['total_calls']} calls")

Export Stats

from agentdna import export_stats

json_str = export_stats(format="json")
csv_str = export_stats(format="csv")

How It Works

1. Decorator wraps your function
2. Each call logs: timestamp, success/failure, latency, input/output sizes, errors
3. Data batches to ~/.agentdna/observe.db (SQLite, WAL mode)
4. Read stats anytime via get_stats() or agentdna stats

No dependencies beyond Python stdlib. click is only needed for the CLI.

Data Location

Default: ~/.agentdna/observe.db

Custom path:

import os
os.environ["AGENTDNA_DB_PATH"] = "/path/to/my/observe.db"

Why AgentDNA?

Without AgentDNA	With AgentDNA
print("done!")	Track every call with latency + errors
Hope nothing breaks	Know exactly when and what breaks
Debug blind	P95 latency, error breakdown, trends
Lose data on restart	SQLite persistence

Installation

pip install agentdna

Optional features:

pip install agentdna[discovery]   # agent discovery (httpx, PyYAML)
pip install agentdna[server]      # run registry server (FastAPI)

License

Apache 2.0