Description
When two agents backed by Anthropic AnthropicClient are connected via WorkflowBuilder.add_edge(), the first agent's output (assistant-role messages) is passed directly as context to the second agent. Anthropic's API rejects this because it requires the conversation to end with a user message.
This means any multi-agent workflow using Anthropic models fails on the second agent invocation. The issue is in AgentExecutor with the default context_mode="full" — it forwards the full conversation history including assistant messages without re-roling them for providers that don't support assistant prefill.
Expected Behavior:
The workflow should handle message role conversion automatically when passing context between agents, regardless of the downstream provider. Either:
AgentExecutor should detect that the target agent's provider doesn't support assistant prefill and re-role assistant messages as user messages, or
The WorkflowBuilder should insert a role-conversion step when connecting agents backed by Anthropic.
Code Sample
#!/usr/bin/env python3
"""
Minimal repro: Agent Framework workflow fails with Anthropic models.
When two agents are connected via WorkflowBuilder, the first agent's output
(assistant-role messages) is passed directly to the second agent. Anthropic's
API rejects this because the conversation must end with a user message.
Error:
anthropic.BadRequestError: 400 - This model does not support assistant
message prefill. The conversation must end with a user message.
Expected: The workflow should automatically handle message role conversion
when the downstream agent uses Anthropic.
"""
from __future__ import annotations
import asyncio
from agent_framework import AgentResponseUpdate, WorkflowBuilder
from agent_framework_anthropic import AnthropicClient
from dotenv import load_dotenv
load_dotenv()
async def main() -> None:
client = AnthropicClient(model_id="claude-sonnet-4-6")
agent_a = client.as_agent(
name="AgentA",
instructions="You are a helpful assistant. Respond briefly.",
)
agent_b = client.as_agent(
name="AgentB",
instructions="You are a helpful assistant. Summarize what you received.",
)
# AgentA's assistant output flows directly to AgentB.
# Anthropic rejects this because the conversation ends with an assistant message.
workflow = (
WorkflowBuilder(start_executor=agent_a)
.add_edge(agent_a, agent_b)
.build()
)
async for event in workflow.run("Say hello", stream=True):
if isinstance(event, AgentResponseUpdate) and event.text:
print(event.text, end="", flush=True)
print()
if __name__ == "__main__":
asyncio.run(main())Error Messages / Stack Traces
anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'This model does not support assistant message prefill. The conversation must end with a user message.'}Package Versions
agent-framework-rc6
Python Version
3.12.1
Additional Context
No response
Description
When two agents backed by Anthropic AnthropicClient are connected via WorkflowBuilder.add_edge(), the first agent's output (assistant-role messages) is passed directly as context to the second agent. Anthropic's API rejects this because it requires the conversation to end with a user message.
This means any multi-agent workflow using Anthropic models fails on the second agent invocation. The issue is in AgentExecutor with the default context_mode="full" — it forwards the full conversation history including assistant messages without re-roling them for providers that don't support assistant prefill.
Expected Behavior:
The workflow should handle message role conversion automatically when passing context between agents, regardless of the downstream provider. Either:
AgentExecutor should detect that the target agent's provider doesn't support assistant prefill and re-role assistant messages as user messages, or
The WorkflowBuilder should insert a role-conversion step when connecting agents backed by Anthropic.
Code Sample
#!/usr/bin/env python3 """ Minimal repro: Agent Framework workflow fails with Anthropic models. When two agents are connected via WorkflowBuilder, the first agent's output (assistant-role messages) is passed directly to the second agent. Anthropic's API rejects this because the conversation must end with a user message. Error: anthropic.BadRequestError: 400 - This model does not support assistant message prefill. The conversation must end with a user message. Expected: The workflow should automatically handle message role conversion when the downstream agent uses Anthropic. """ from __future__ import annotations import asyncio from agent_framework import AgentResponseUpdate, WorkflowBuilder from agent_framework_anthropic import AnthropicClient from dotenv import load_dotenv load_dotenv() async def main() -> None: client = AnthropicClient(model_id="claude-sonnet-4-6") agent_a = client.as_agent( name="AgentA", instructions="You are a helpful assistant. Respond briefly.", ) agent_b = client.as_agent( name="AgentB", instructions="You are a helpful assistant. Summarize what you received.", ) # AgentA's assistant output flows directly to AgentB. # Anthropic rejects this because the conversation ends with an assistant message. workflow = ( WorkflowBuilder(start_executor=agent_a) .add_edge(agent_a, agent_b) .build() ) async for event in workflow.run("Say hello", stream=True): if isinstance(event, AgentResponseUpdate) and event.text: print(event.text, end="", flush=True) print() if __name__ == "__main__": asyncio.run(main())Error Messages / Stack Traces
anthropic.BadRequestError: Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'This model does not support assistant message prefill. The conversation must end with a user message.'}Package Versions
agent-framework-rc6
Python Version
3.12.1
Additional Context
No response