How to wire agent.generate() with workflow suspend() in-chat approvals?
AgentMemory
Hey everyone 
looking for some feedback on our Mastra setup
We’re building an internal chat support agent where an operator types something like:
update field X to value Y on record 123
What we want:
• Agent understands intent and routes to the right domain
• If something’s missing, it asks follow-up questions in chat
• Once it has everything, it triggers a Mastra workflow
• If the workflow suspend()s (for approval), the agent shows the changes in chat
• Operator types approve / decline → agent resumes the workflow
• Workflow finishes → result is posted back in chat
Basically: one conversational agent that both gathers data and executes workflows — fully inside chat.
Things we’re unsure about:
Workflow + agent.generate()
If a workflow suspends mid-run:
Does the agent receive the suspend payload and reason about it? Or does generate() block until completion?
Resume pattern
After a suspend, the next user message should resume the workflow.
Should we store runId + stepId and expose a resumeWorkflow tool? Or is there a more native Mastra pattern for this?
Router → domain agents (~10 domains)
- What’s the cleanest approach?
- Flatten all tools into one big agent?
-Agent-as-tool pattern?
-Native handoff/delegation?
Multiple suspends
Our workflows suspend twice (review → confirm).
Does Mastra handle multi-suspend conversational flows cleanly, or do we need to orchestrate state ourselves?
Long conversations (10+ turns)
Is basic Mastra Memory (lastMessages: N) enough, or is there a better pattern for durable workflow-driven chats?
If anyone has implemented agent + workflow + suspend/resume fully inside chat, would really appreciate guidance
looking for some feedback on our Mastra setupWe’re building an internal chat support agent where an operator types something like:
update field X to value Y on record 123
What we want:
• Agent understands intent and routes to the right domain
• If something’s missing, it asks follow-up questions in chat
• Once it has everything, it triggers a Mastra workflow
• If the workflow suspend()s (for approval), the agent shows the changes in chat
• Operator types approve / decline → agent resumes the workflow
• Workflow finishes → result is posted back in chat
Basically: one conversational agent that both gathers data and executes workflows — fully inside chat.
Things we’re unsure about:
Workflow + agent.generate()If a workflow suspends mid-run:
Does the agent receive the suspend payload and reason about it? Or does generate() block until completion?
Resume patternAfter a suspend, the next user message should resume the workflow.
Should we store runId + stepId and expose a resumeWorkflow tool? Or is there a more native Mastra pattern for this?
Router → domain agents (~10 domains)- What’s the cleanest approach?
- Flatten all tools into one big agent?
-Agent-as-tool pattern?
-Native handoff/delegation?
Multiple suspendsOur workflows suspend twice (review → confirm).
Does Mastra handle multi-suspend conversational flows cleanly, or do we need to orchestrate state ourselves?
Long conversations (10+ turns)Is basic Mastra Memory (lastMessages: N) enough, or is there a better pattern for durable workflow-driven chats?
If anyone has implemented agent + workflow + suspend/resume fully inside chat, would really appreciate guidance
