AgentOS Tools

AgentOSTools gives an agent a read-only ops view of the AgentOS it runs on: usage, latency, failures, tool statistics, schedules, eval history, pending approvals and runtime-built components. The toolkit takes the database, never the AgentOS instance — agents are constructed before the OS, so every tool reads only from db.

Files

File	What it teaches
`platform_ops_agent.py`	Generate traced activity with a worker agent, then answer platform questions with an ops agent using AgentOSTools.

Prerequisites

bash

./scripts/demo_setup.sh
export OPENAI_API_KEY=...

Run platform_ops_agent.py

bash

.venvs/demo/bin/python cookbook/05_agent_os/25_agentos_tools/platform_ops_agent.py

The demo runs the worker agent twice, verifies the activity is visible through the toolkit (run counts, tool statistics, session metrics), then asks the ops agent for a platform summary.

Notes

No tool mutates platform state; schedule, approval and component management are deliberately not exposed. The one write is the metrics rollup refresh inside get_platform_metrics (derived data, no user content). The tools read the database directly, so AgentOS endpoint scopes do not apply — expose the ops agent to operators, and trim surfaces with the enable flags for wider audiences.
Sensitive payloads are never returned: span attributes, approval tool arguments and schedule run input/output can hold conversation content and are excluded from every tool result.
Per-surface flags (metrics, traces, schedules, evals, components, approvals) control which tools a deployment exposes.
On SQLite, p95_duration_ms is reported as null (no SQL percentile support); PostgreSQL reports real percentiles. The payload says so when it applies.