Repository navigation
Python: Add context_mode to control what gets sent to agents in workflows #3688
Description
Summary
Investigate adding a context_mode option to control what context gets passed between agents in a workflow.
Background
From a user question:
My scenario involves a deterministic sequential flow with multiple agents, so I'm using the Agent framework's sequential workflow. In my case, I don't need to pass the whole context to each agent in the workflow; I only want each agent to receive the final output from the previous one as input. From what I understand, the default behavior is to pass the entire context. How can I adjust this?
Current Workaround
Today, users need to create a custom executor that transforms the conversation between agents:
class LastOutputOnlyExecutor(Executor):
@handler
async def transform(
self,
messages: list[ChatMessage],
ctx: WorkflowContext[list[ChatMessage]]
) -> None:
# Extract only the last assistant message
last_assistant = [m for m in messages if m.role == "assistant"][-1:]
await ctx.send_message(last_assistant)Then build the flow like:
workflow = SequentialBuilder().participants([
agent1,
LastOutputOnlyExecutor(id="filter-1"),
agent2,
LastOutputOnlyExecutor(id="filter-2"),
agent3,
]).build()Proposed Solution
Investigate adding a context_mode parameter to reduce verbosity. Possible modes:
full(default) - pass entire conversation contextlast_output- pass only the last assistant messagelast_n- pass last N messages- Custom filter function
Scope
This feature should be evaluated for all orchestration patterns, not just sequential:
- Sequential workflows
- Concurrent workflows
- Group chat
- Other orchestration patterns
Activity
[-]Add context_mode to control what gets sent to agents in workflows[/-][+]Python: Add context_mode to control what gets sent to agents in workflows[/+]moonbox3 commentedon Mar 5, 2026
See #4038 for further ideas, if not represented here.
droideronline commentedon Mar 6, 2026
Thanks for covering these ideas - really appreciate the thorough exploration of
context_modeoptions. One thing I'd like to highlight again: once this feature is implemented, it will also need to be supported in declarative workflows (e.g., YAML-based definitions). The goal is to allow users to configure context passing behavior — such asfull,last_output, orlast_n— directly in the workflow definition, without requiring any code changes. This is especially important for scenarios where non-developer users or ops teams need to adjust workflow behavior through configuration alone.TaoChenOSU commentedon Mar 6, 2026
When Agent A invokes Agent B, we also need to consider the messages that are already in Agent B's cache. Currently, we are replacing the cache of Agent B with Agent A's context, either the full conversation or the response: https://github.com/microsoft/agent-framework/blob/main/python/packages/core/agent_framework/_workflows/_agent_executor.py#L150
I think we should also have a mode to control if an agent should retain its own context when receiving a request from another agent.
TaoChenOSU commentedon Mar 6, 2026
Perhaps a mode on the receiving agent executor that says:
Note that if not handled carefully, retaining cache and supplying full context from the previous agent may cause unexpected behavior.