Foundry YAML Orchestrator

Run Microsoft Agent Framework teams directly from YAML. There is no YAML-to-Python generation step: one root orchestrator.py loads a team definition, loads each referenced agent definition in order, binds Python tools, and runs the selected orchestration pattern.

python -m pip install -r requirements.txt
az account show --query "{tenant:tenantId, subscription:name}" -o table

If Azure CLI is not authenticated, sign in to the tenant that owns the Foundry project before running a team.

Billing:

python orchestrator.py --team-yaml agent_teams/vf_billing_team/team.yaml

The configured task runs as the first turn. The process then displays You> and keeps the same team sessions open for follow-up messages. Type exit or quit to close the chat.

Triage:

python orchestrator.py --team-yaml agent_teams/vf_triage_team/team.yaml

Override the first chat turn without editing YAML:

python orchestrator.py `
  --team-yaml agent_teams/vf_billing_team/team.yaml `
  --task '{"userQuery":"Show my current bill","selectedAccountReference":"personal"}'

Run exactly one task and exit for scripts or smoke tests:

python orchestrator.py `
  --team-yaml agent_teams/vf_billing_team/team.yaml `
  --once

After installing the project, the equivalent console command is:

foundry-team --team-yaml agent_teams/vf_billing_team/team.yaml

Each team directory contains:

team.yaml               Orchestration pattern, ordered agents, default task
.env.example             Foundry endpoint and model settings
agents/*.yaml            Agent model, instructions, and tool declarations
src/tools.py             Python implementations exposed through TOOL_REGISTRY

Each subdirectory under agent_teams/ is a source-controlled domain team, such as billing, roaming, or triage.

A team file declares its runtime files and orchestration:

name: example-team
runtime:
  env_file: ./.env
  tools_file: ./src/tools.py
orchestration:
  pattern: sequential
  chain_only_agent_responses: true
  agents:
    - ./agents/first_agent.yaml
    - ./agents/second_agent.yaml
  task: Process the request.

For sequential, agents execute in the listed order. With chain_only_agent_responses: true, each agent receives only the immediately preceding agent response. Agent prompts should therefore define a stable output contract for the next stage.

Function tools declared in agent YAML must have matching entries in TOOL_REGISTRY. MCP tools are constructed directly from their YAML declarations.

Supported patterns are sequential, concurrent, group_chat, and handoff.

Copy the team example file and set its Foundry endpoint:

Copy-Item agent_teams/vf_billing_team/.env.example agent_teams/vf_billing_team/.env

The runtime loads the .env adjacent to team.yaml. Real .env files are ignored by Git.

python -m pytest tests/test_orchestrator.py -q

See ARCHITECTURE.md for runtime internals and USAGE_GUIDE.md for adding a team.