ADK 2.0.0 GA: task-mode LlmAgents disabled as workflow graph nodes (regression from 2.0.0b1)

[markns]

Summary

ADK 2.0.0 GA disabled the documented "task agent as workflow graph node" pattern via a new validator in Workflow. The validator raises:

Agent '<name>' has mode='task' and cannot be used as a workflow graph node. Use a chat coordinator with task sub-agents, or dispatch dynamically via ctx.run_node from a function node.

This breaks code that worked under 2.0.0b1 and uses the canonical agent_in_workflow pattern shipped in upstream samples (contributing/task_samples/agent_in_workflow/). It is not flagged as a breaking change in CHANGELOG-v2.md — line 14 still advertises "task agents as workflow nodes" as a 2.0 feature.

The validator's own docstring and the xfail(strict=True) markers attached to previously-passing tests describe the gate as temporary, pending a scheduler fix. Filing this so external users have a tracking handle for re-enablement.

Source

src/google/adk/workflow/_workflow.py:192-219 — _validate_no_task_mode_graph_nodes:

def _validate_no_task_mode_graph_nodes(self) -> None:
    """Reject ``mode='task'`` LlmAgents that appear as static graph nodes.

    Task-mode agents are multi-turn — they pause for user replies and
    expect the original ``node_input`` (the task brief) to remain visible
    across re-dispatches.  The workflow scheduler currently overwrites
    ``node_input`` with the latest user message on every re-entry, so the
    task brief is lost and the agent loses context.  Until the scheduler
    preserves the originating ``node_input`` on resume, task agents may
    only be used:

      * as chat sub-agents of an LlmAgent coordinator (FC delegation
        via ``_TaskAgentTool`` / ``_dispatch_task_fc``), or
      * dispatched dynamically via ``ctx.run_node`` from a custom
        function node — never as static workflow graph nodes.
    """

Introduced in commit 16227935 ("chore: switch main to v2.0.0 GA (transition to v2)", 2026-05-19), which appears to be a squash of the GA prep work — no granular PR is visible in the public history.

Upstream xfail evidence

These previously-passing tests were marked xfail(strict=True) in the same commit with the message:

"mode='task' workflow graph nodes temporarily disabled; re-enable when scheduler preserves originating node_input on resume."

• tests/unittests/workflow/test_llm_agent_as_node.py — at least lines 261, 320, 384
• tests/unittests/workflow/test_workflow_live.py — lines 104, 247, 331, 557
• tests/unittests/workflow/test_workflow_llm_agent_interruptions.py — lines 130, 421

strict=True means these are upstream's own admission that the prior behavior worked and is now intentionally blocked — not flaky tests being skipped.

Reproduction

from google.adk.agents import Agent
from google.adk import Workflow

task_agent = Agent(name="my_task", mode="task", instruction="...")

# 2.0.0b1: constructs fine.
# 2.0.0 GA: raises ValueError at construction time.
wf = Workflow(name="root", edges=[("START", task_agent)])

Environment:

• google-adk==2.0.0
• Python 3.14
• macOS / Linux (platform-independent — fails at module import)

Impact

• Anyone using the agent_in_workflow pattern documented in contributing/task_samples/agent_in_workflow/ cannot upgrade from 2.0.0b1 → 2.0.0 without restructuring.
• The two supported workarounds (chat coordinator with task sub-agents via _TaskAgentTool/_dispatch_task_fc, or dynamic dispatch via ctx.run_node from a function node) both work but have meaningfully different routing semantics from the static-edges shape, so this is a real refactor, not a one-line swap.
• The CHANGELOG should call this out as a breaking change for 2.0.0b1 → 2.0.0 until the gate is lifted.

Tracking ask

1. Confirm the gate is temporary as the docstring states.
2. Track the scheduler fix (preserving originating node_input on resume) here or link the canonical tracking issue.
3. Add a CHANGELOG entry for the GA breaking change so users upgrading from 2.0.0b1 aren't surprised.

Related issues

• ctx.run_node() sequential calls to multiple single_turn agents — second agent receives empty/ignored node_input #5686 — ctx.run_node() sequential calls to multiple single_turn agents — second agent receives empty/ignored node_input (same node_input-handling family, dynamic-dispatch surface)
• V2.0 beta - Contents lost between turns within a dynamic workflow #5461 (closed) — node_input lost across tool rounds within a single run_node invocation (same family, inner-LlmAgent surface)
• Raw event.content assigned to child.output without transformation in _reconstruct_node_states during node re-run, for the nodes with message_as_input = True #5553 — Raw event.content assigned to child.output without transformation in _reconstruct_node_states during node re-run (adjacent re-run/resume area)

[Jacksunwei]

Hi @markns , thanks for reporting! We find an issue of using it in workflow and we're fixing it soon and reenable it.

[markns]

Thanks @Jacksunwei - I'll keep an eye out for the fix! In general, I think it would be great if there were some more guidance about how graph workflows and agents (are intended to) interact. There's some discussions open also (#5581, #5671)