Description
HandoffAgentExecutor.__init__ folds the configured autonomous turn limit through truthiness (python/packages/orchestrations/agent_framework_orchestrations/_handoff.py:282):
self._autonomous_mode_turn_limit = autonomous_mode_turn_limit or _DEFAULT_AUTONOMOUS_TURN_LIMIT
with _DEFAULT_AUTONOMOUS_TURN_LIMIT = 50 (:229), so an explicitly configured 0 is silently replaced by 50. 0 is the natural way for a caller to say "this agent must not continue on its own", and the public builder path makes it reachable without constructing the executor directly: HandoffBuilder.with_autonomous_mode(turn_limits={agent_id: 0}) stores the mapping and build() hands it to the executor unchanged through self._autonomous_mode_turn_limits.get(id, None) (:1210), where the or above then swallows it.
The consumer treats the value as a plain counter bound (:480):
if self._autonomous_mode and self._autonomous_mode_turns < self._autonomous_mode_turn_limit:
With the intended limit of 0 that branch is never taken, so the executor emits request_info as soon as the agent answers without requesting a handoff. With the swallowed 50, the executor instead injects the autonomous prompt and re-runs the agent up to 50 more times first. The docstring describes the default as belonging to the unset case — "Maximum number of autonomous turns before returning control to the user. If not provided, a default turn limit will be used." (:934) — and "not provided" is None, not 0.
The cost is not cosmetic: a caller who asks one agent to stop autonomously gets up to 50 unwanted model calls and 50 unwanted replies before control returns.
Code Sample
Reproduces with the module's existing mock harness, no API keys needed. Added to python/packages/orchestrations/tests/test_handoff.py, right after test_autonomous_mode_resumes_user_input_on_turn_limit:
async def test_autonomous_mode_zero_turn_limit_returns_control_immediately():
"""A turn limit of 0 means no autonomous continuation, not the default turn limit."""
triage_client = MockChatClient(name="triage", handoff_to="worker")
worker_client = MockChatClient(name="worker")
triage: Agent[Any] = Agent(id="triage", name="triage", client=triage_client,
require_per_service_call_history_persistence=True)
worker: Agent[Any] = Agent(id="worker", name="worker", client=worker_client,
require_per_service_call_history_persistence=True)
workflow = (
HandoffBuilder(participants=_as_handoff_agents(triage, worker), termination_condition=lambda _: False)
.with_start_agent(_as_handoff_agent(triage))
.with_autonomous_mode(agents=[worker], turn_limits={resolve_agent_id(worker): 0})
.build()
)
events = await _drain(workflow.run("Start", stream=True))
assert len(worker_client.received_messages) == 1, "A zero turn limit must not grant autonomous turns"
Error Messages / Stack Traces
No exception is raised — the agent just keeps going. The assertion above fails on main:
> assert len(worker_client.received_messages) == 1, "A zero turn limit must not grant autonomous turns"
E AssertionError: A zero turn limit must not grant autonomous turns
E assert 51 == 1
packages\orchestrations\tests\test_handoff.py:1358
51 = the one reply the worker owes the handoff plus the 50 autonomous turns the configured 0 was supposed to forbid.
Package Versions
agent-framework-core: 1.19.0, agent-framework-orchestrations: 1.2.0 (source main @ a91fe668)
Additional Context
HandoffAgentExecutor is the only consumer of turn_limits, and the same truthiness pattern is not used for the other autonomous settings (prompts maps to prompts or {}, which is correct for a container), so replacing the or with an is None check is the whole fix.
#6328 lists "Turn limit = 0 (default limit of 50 applied)" as one of its observations while investigating a separate 49-vs-50 boundary question with live models. This issue isolates only the 0-becomes-default part, which is deterministic and unit-testable; the boundary question in #6328 is unaffected by this and stays open.
Description
HandoffAgentExecutor.__init__folds the configured autonomous turn limit through truthiness (python/packages/orchestrations/agent_framework_orchestrations/_handoff.py:282):with
_DEFAULT_AUTONOMOUS_TURN_LIMIT = 50(:229), so an explicitly configured0is silently replaced by 50.0is the natural way for a caller to say "this agent must not continue on its own", and the public builder path makes it reachable without constructing the executor directly:HandoffBuilder.with_autonomous_mode(turn_limits={agent_id: 0})stores the mapping andbuild()hands it to the executor unchanged throughself._autonomous_mode_turn_limits.get(id, None)(:1210), where theorabove then swallows it.The consumer treats the value as a plain counter bound (
:480):With the intended limit of
0that branch is never taken, so the executor emitsrequest_infoas soon as the agent answers without requesting a handoff. With the swallowed50, the executor instead injects the autonomous prompt and re-runs the agent up to 50 more times first. The docstring describes the default as belonging to the unset case — "Maximum number of autonomous turns before returning control to the user. If not provided, a default turn limit will be used." (:934) — and "not provided" isNone, not0.The cost is not cosmetic: a caller who asks one agent to stop autonomously gets up to 50 unwanted model calls and 50 unwanted replies before control returns.
Code Sample
Reproduces with the module's existing mock harness, no API keys needed. Added to
python/packages/orchestrations/tests/test_handoff.py, right aftertest_autonomous_mode_resumes_user_input_on_turn_limit:Error Messages / Stack Traces
No exception is raised — the agent just keeps going. The assertion above fails on
main:51 = the one reply the worker owes the handoff plus the 50 autonomous turns the configured
0was supposed to forbid.Package Versions
agent-framework-core: 1.19.0, agent-framework-orchestrations: 1.2.0 (source
main@a91fe668)Additional Context
HandoffAgentExecutoris the only consumer ofturn_limits, and the same truthiness pattern is not used for the other autonomous settings (promptsmaps toprompts or {}, which is correct for a container), so replacing theorwith anis Nonecheck is the whole fix.#6328 lists "Turn limit = 0 (default limit of 50 applied)" as one of its observations while investigating a separate 49-vs-50 boundary question with live models. This issue isolates only the
0-becomes-default part, which is deterministic and unit-testable; the boundary question in #6328 is unaffected by this and stays open.