driftbase 0.1.1

Behavioral drift monitoring for AI agents.

Driftbase

Behavioral drift monitoring for AI agents: track runs locally, compare versions, get a drift score.

When you ship a new prompt, model, or tool set, agent behavior can change in subtle ways. Driftbase records each run (tool names, latency, outcome) in a local SQLite DB and lets you diff any two versions so you see a numeric drift score and per-dimension breakdown before or after deploy.

---

Quickstart

1. Install

pip install driftbase[local]

2. Add the decorator

# my_agent.py
from driftbase import track

@track(version="v1.0", environment="production")
def my_agent(user_input: str) -> str:
    # your agent logic
    return "done"

3. Run your agent

Generate some runs for at least two versions (e.g. change code and use version="v2.0" for new runs).

python -c "from my_agent import my_agent; my_agent('hello')"
# Run a few more times, then switch to v2.0 and run again.

4. Run diff

driftbase diff v1.0 v2.0

5. See output

You get a threshold panel, a metrics table, tool frequency diff, optional sequence shifts, and a root-cause hypothesis. Example (with comments indicating where rich applies color):

# Panel: red border if above threshold, green if within
┌─ ▲ ABOVE THRESHOLD ─────────────────────────────────────────────────┐
│ Drift score 0.34 is above threshold 0.20. Consider investigating...   │
└──────────────────────────────────────────────────────────────────────┘

# Table: Drift — v1.0 → v2.0
┌─────────────────┬──────────┬─────────┬────────┐
│ Metric          │ Baseline │ Current │ Delta  │
├─────────────────┼──────────┼─────────┼────────┤
│ Overall drift   │     0.00 │    0.34 │  +0.34 │  # red if ≥ threshold
│ Decision drift  │     0.00 │    0.22 │  +0.22 │
│ Latency drift   │     0.00 │    0.18 │  +0.18 │
│ Error drift     │     0.00 │    0.00 │  +0.00 │
└─────────────────┴──────────┴─────────┴────────┘

# Tool call frequency diff (top 20 tools, Δ % in green/red/dim)
# Optional: Top 3 sequence shifts, Root cause hypothesis panel
# Footer: Runs: v1.0 (n=50) → v2.0 (n=50) · No data left your machine

---

How it works

Runs are written to SQLite in a background thread so your app is not blocked. When you run driftbase diff, the CLI loads runs for the two versions, builds a behavioral fingerprint for each (tool distributions, latency percentiles, error rate), and computes a divergence score between them. The score and per-dimension deltas tell you how much behavior changed.

---

Privacy

• Captured and stored locally: Tool call names and order, latency, token counts, error/retry counts, outcome label (e.g. resolved/error). No raw user or model content.
• Hashed then discarded: A hash of the task input and a hash of the output structure are stored; the original text is not.
• Never stored or read: Raw user messages, raw agent output, system prompts, API keys, user identifiers.

Use driftbase inspect --run last to see the exact breakdown for any run.

---

CLI reference

Command	Description	Example
versions	List deployment versions and run counts	driftbase versions
diff	Compare two versions; optional last N vs baseline	driftbase diff v1.0 v2.0 or driftbase diff v1.0 local --last 20
inspect	Show what was captured/dropped for a run	driftbase inspect --run last
report	Generate markdown/JSON/HTML drift report	driftbase report v1.0 v2.0 -o report.md
watch	Live drift monitor against a baseline	driftbase watch --against v1.0
push	Send local runs to Driftbase platform API	driftbase push (uses DRIFTBASE_API_URL, DRIFTBASE_API_KEY)

---

Frameworks supported

The @track() decorator auto-detects and captures from:

• LangChain — tool calls via callbacks
• LangGraph — same as LangChain
• LlamaIndex — function_call and callback events
• OpenAI — chat.completions.create tool_calls and usage
• Generic — any callable; times the call and optionally parses tool_calls from the return value

---

Docs and product

• Documentation
• Driftbase