Run CompanyOS Products, Tasks and a Product Manager inside AgentOS

Agno's AgentOS(db=db, company=CompanyConfig()) turns on CompanyOS and mounts its routes under /company, with tables created on the first write. Leave company out and AgentOS adds no Company route, table or permission scope.

from agno.agent import Agent
from agno.db.sqlite import SqliteDb
from agno.models.openai import OpenAIResponses
from agno.os import AgentOS, CompanyConfig
 
db = SqliteDb(db_file="company.db")
writer = Agent(id="docs-writer", model=OpenAIResponses(id="gpt-5.6-luna"), db=db)
 
agent_os = AgentOS(
    db=db,
    agents=[writer],
    company=CompanyConfig(model=OpenAIResponses(id="gpt-5.6-luna")),
)
app = agent_os.get_app()

A Product is a named area of responsibility with a purpose. Projects inside it move from proposed to active to done. Tasks start in backlog and move through ready, in_progress and in_review to done, and either record can be cancelled before it finishes. Every move goes through POST /company/tasks/{id}/transitions (or the Projects twin) with a target and a reason, and the record keeps that history. A person makes a Task Ready by naming a registered agent, team or workflow as its executor, and AgentOS checks that the caller may run it. From then on the Task's scope is frozen until someone returns it to Backlog.

Each Product also gets a playbook wiki with brief.md for human direction and status.md for current work, and the company shares one more at /company/wiki. A write must carry the version it read, so a stale edit gets a 409. Decisions go to /company/decisions as an append-only log.

The Product Manager is a reserved agent, product-manager-internal, served at /agents/product-manager-internal/runs or as an OS Manager member when OS Manager is on. Its tools read Products, Projects, Tasks and the wiki, propose and edit Backlog work, update status.md and log decisions. None of them can accept or finish a Task, change an owner or schedule a run, and AgentOS injects the caller's identity outside the model's tool arguments.

We drove the whole loop over HTTP on SQLite. A scripted stand-in model, built like the offline planner in the release's own tests, played the Product Manager's model.

CompanyManager and AsyncCompanyManager expose the same records from Python, and CompanyTools can be published to MCP clients with MCPConfig(tools=[...]).

See the cookbook, and learn more about AgentOS in the documentation.