v2.8.5

New AgentOSTools gives an agent a read-only ops view of the AgentOS it runs on. Point it at your database and it can report on usage, latency, failures, schedules, evals, components, and pending approvals, then answer questions about any of it. Things like "which tool was slowest today," "how many runs failed this week," or "what's waiting on approval." You get answers grounded in real traces instead of clicking through a dashboard to assemble them yourself.

from agno.agent import Agent
from agno.db.sqlite import SqliteDb
from agno.tools.agentos import AgentOSTools

db = SqliteDb(db_file="tmp/platform.db")

ops_agent = Agent(
    tools=[AgentOSTools(db=db)],
    instructions="Answer ops questions about this AgentOS. Ground every number in a tool result.",
    db=db,
)

ops_agent.print_response("Which tool was slowest today, and how many runs failed?")

‍

The toolkit reads from the database, not from a live AgentOS handle. Your agents get built before the OS does, so wiring the tools to the db is what keeps this cleanly read-only: an ops agent can inspect the whole platform without changing a single agent, session, or config. It watches, it doesn't operate. It reports on whatever your deployment has traced, so with tracing on, the ops agent always has something real to pull from.

There's something usefully recursive here. The same framework you use to build agents now runs the agent that keeps an eye on them. Rather than a human refreshing a metrics page, you can put an agent on call that reads the traces, notices the slow tool or the jump in failures, and tells you in a sentence what's going on.

View the platform ops agent cookbook for more.