Praison AI

Container Default LLM

One `llm=` on AgentTeam or AgentFlow fills in members that didn't pick a model — pinned agents keep theirs.

One model= on AgentTeam or AgentFlow becomes the default model for every member that never named one, while any agent that pinned its own model keeps it.

Diagram

Container Default LLM

No

Yes

⚙️ Container llm=

🔍 Member named
a model?

✅ Fill in

🔒 Keep member’s model

Quick Start

  1. One team, one model
    python
    from praisonaiagents import Agent, AgentTeam

researcher = Agent(name="Researcher", instructions="Research.")
writer     = Agent(name="Writer",     instructions="Write.")

team = AgentTeam(
    agents=[researcher, writer],
    model="anthropic/claude-sonnet-4-5",
)
team.start("Summarise the latest AI safety research")

    The team’s model= flows to every member that didn’t pick its own.

  2. Same rule for a flow
    python
    from praisonaiagents import Agent
from praisonaiagents.workflows import AgentFlow

researcher = Agent(name="Researcher", instructions="Research.")
writer     = Agent(name="Writer",     instructions="Write.")

flow = AgentFlow(
    steps=[researcher, writer],
    model="anthropic/claude-sonnet-4-5",
)
flow.start("Summarise the latest AI safety research")

    AgentFlow(model=…) fills in ready-made steps the same way AgentTeam does.

How It Works

The container’s llm= is a default, not an override. A member that named its own llm= / model= / auth= keeps it; a member with no model adopts the container’s.

Diagram

Yes

No, and container has llm=

No, and no container llm=

Did the Agent name
its own llm/model/auth?

Keep the agent’s model

Adopt container llm=

Agent’s env-aware default
OPENAI_MODEL_NAME or gpt-4o-mini

A pinned member keeps its model even when the container names another:

python
from praisonaiagents import Agent, AgentTeam

pinned = Agent(name="Reasoner",   instructions="Think hard.", llm="gpt-4o")
cheap  = Agent(name="Summariser", instructions="Summarise.")

team = AgentTeam(
    agents=[pinned, cheap],
    llm="anthropic/claude-sonnet-4-5",
)
# pinned.llm == "gpt-4o"                       (kept)
# cheap.llm  == "anthropic/claude-sonnet-4-5"  (filled in)

manager_llm Is Separate

manager_llm= is the hierarchical manager’s own model and never touches the members.

python
from praisonaiagents import Agent, AgentTeam

member = Agent(name="Worker", instructions="Do the work.")

team = AgentTeam(
    agents=[member],
    llm="anthropic/claude-sonnet-4-5",  # member default
    manager_llm="gpt-4o",               # hierarchical manager only
)
# team.manager_llm == "gpt-4o"                       (manager keeps its own model)
# member.llm       == "anthropic/claude-sonnet-4-5"  (member fills in)

AgentFlow’s Fallback Is Not Propagated

AgentFlow falls back to "gpt-4o-mini" internally when it needs a model to run, but that fallback is not a caller’s choice — so it is never pushed onto steps.

python
from praisonaiagents import Agent
from praisonaiagents.workflows import AgentFlow

researcher = Agent(name="Researcher", instructions="Research.")

# No container llm= — each step keeps its own env-aware default,
# NOT the flow's "gpt-4o-mini" run-time fallback.
flow = AgentFlow(steps=[researcher])
flow.start("Summarise the latest AI safety research")

An AgentFlow(steps=[…]) with no llm= leaves each step’s model to Agent‘s own env-aware default (OPENAI_MODEL_NAME, else gpt-4o-mini).

Common Patterns

Pin one member, let the rest follow the container — no code change needed to see the split.

python
from praisonaiagents import Agent, AgentTeam

reasoner = Agent(name="Reasoner",   instructions="Think hard.", llm="gpt-4o")
writer   = Agent(name="Writer",     instructions="Write.")
editor   = Agent(name="Editor",     instructions="Edit.")

team = AgentTeam(
    agents=[reasoner, writer, editor],
    llm="anthropic/claude-sonnet-4-5",
)
team.start("Draft and polish a launch post")
# reasoner calls gpt-4o; writer and editor call anthropic/claude-sonnet-4-5
Panel agents are skipped

A panel descriptor (panel:<name>) is not a model name, so the fill-in leaves it alone.

 `auth=` counts as pinned

A subscription seat pins the vendor, so a team default of another vendor never overrides it.

 List or dict rosters

AgentTeam accepts a list or a dict (multi-group teams); both forms are filled in.

 Non-Agent members skipped

Members without _apply_default_llm are duck-typed and left untouched.

A member used before the container fills it in caches a dispatcher bound to the old model; the fill-in drops that cache so the next call rebuilds against the new model. Pinned members keep their dispatcher.

Configuration Options

ParamContainerBehaviour
llmAgentTeam, AgentFlowDefault model for members that never named one. A member with its own llm= / model= / auth= keeps it.
manager_llmAgentTeamHierarchical manager’s own model. Never propagated to members.

Best Practices

Set the vendor once at the container level

Put llm= on AgentTeam / AgentFlow and pin per-agent only where the job justifies a different model.

Give the manager its own model

Use manager_llm= for the hierarchical manager — don’t try to reuse the team default there.

Prefer container llm= for whole-team agreement

When every member should share a vendor, one container llm= beats repeating llm= on each Agent.

Upgrade if llm= did nothing before

Earlier versions stored the container llm= but never read it. Upgrade to the version that ships this fill-in so the parameter actually reaches members.

Related

AgentTeam Params

Which params apply at the team level.

 Parameter Scope

Team-level vs per-agent parameter scope.

 AgentTeam

Multi-agent teams overview.

 AgentFlow

Deterministic step pipelines.