I kept seeing the same problem — agents get stuck calling the same tool 50 times, wander off-task, or burn through token budgets before anyone notices. The big observability platforms exist but they're heavy for solo devs and small teams.

So I built DriftShield Mini — a lightweight Python library that wraps your existing LangChain/CrewAI agent, learns what "normal" looks like, and fires Slack/Discord alerts when something drifts.

3 detectors: - Action loops (repeated tool calls, A→B→A→B cycles) - Goal drift (agent wandering from its objective, using local embeddings) - Resource spikes (abnormal token/time usage vs baseline)

4 lines to integrate:

    from driftshield import DriftMonitor
    monitor = DriftMonitor(agent_id="my-agent", alert_webhook="https://hooks.slack.com/...")
    agent = monitor.wrap(existing_agent)
    result = agent.invoke({"input": "your task"})

100% local — SQLite + CPU embeddings. Nothing leaves your machine except the alerts you configure.

pip install driftshield-mini GitHub: https://github.com/ThirumaranAsokan/Driftshield-mini

v0.1 — built this solo. Would genuinely love feedback on what agent reliability problems you're hitting. What should I build next?