Agent Security 101: stop sending your transactional data to telemetry services (www.ashpreetbedi.com)

Stop sending transactional agent traces (conversation history, tool calls, memory updates, internal reasoning) to third‑party telemetry services — they’re not logs and they shouldn’t leave your infrastructure. The piece argues transactional data is the system’s source of truth (often containing PII, proprietary logic, or sensitive files) and must be persisted for reuse, debugging, auditing, and compliance. Telemetry vendors are built for metrics (latency, token usage, error rates) and cold storage, not for replayable, sensitive traces; shipping full traces creates redundant copies, increases attack surface, complicates joins/linking, and drives egress/retention costs. Practical advice: give each agent a database (Postgres+PgVector for production, SQLite for demos), persist every transaction (inputs, outputs, context, tool calls, memory/knowledge updates), keep everything inside your VPC, and surface the data with an internal UI (Streamlit or AgentOS) rather than an external telemetry dashboard. This architecture preserves least‑privilege, enables reproducibility and auditing, avoids costly data duplication, and lets you apply vector retrieval or other app-level logic directly. The author highlights Agno as an example framework that ships agents with integrated DBs and tooling, but stresses the underlying principle is universal: own your data, not the telemetry vendor.