Another exciting week at Mastra! We're excited to announce vNext of Agent Network and workflow improvements that our users have been asking for.
Let's dive into this week's updates.
Agent Network (vNext)
Mastra's AgentNetwork is an agent and workflow orchestration layer -- in other words, a multi-agent primitive. You pass it a set of agents and workflows, and it decides which ones to call in which order with what inputs. We shipped an initial version back in March, but have iterated a lot since then. Now, it uses Mastra workflows, and we've seen performance improvement on routing, and have additional. We're removing the experimental tag -- go ahead and use it!
Here's a quick demo:
In terms of when to use these, there's a spectrum from workflows (deterministic but less powerful) to agents running in a loop (more powerful, less deterministic) to multi-agent primitives (even more powerful, even less deterministic).
Some features of the new Agent Network:
- Converting unstructured inputs: Turning natural language into structured tasks
- Dynamic collaboration: You can pass a set of agents with different memories, instructions, and tool calling capabilities
- Context-aware routing: Making decisions based on conversation history and memory
Two powerful methods for different use cases:
1// Single task execution
2const result = await network.generate("Tell me about Tokyo", {
3 runtimeContext,
4});
5
6// Complex multi-step tasks - orchestrates multiple primitives
7const result = await network.loop(
8 "Research the top 3 cities in Japan, get their weather, and write a travel guide",
9 { runtimeContext },
10);Key improvements in vNext:
- New
.generate()method: Reflects API of the underlying agent APIs - Improved routing: Better primitive selection based on descriptions and input schemas
- Memory-driven intelligence: Uses conversation context for better routing decisions
Agent Network accepts any Mastra primitive - agents, workflows, and tools - creating unprecedented flexibility for building intelligent AI orchestration.
Get started with complete examples in our repository.
Workflows
Workflow Cancellation Support
We've added a highly requested feature: the ability to programmatically cancel workflow runs.
1// Cancel a running workflow
2await workflowRun.cancel();This new .cancel() method enables workflows to be stopped mid-execution, giving you better control over long-running processes and resource management #5534.
Memory
Custom Language Models for Title Generation
You can now specify a custom language model when generating titles in the Memory module, allowing you to override the default model set on the agent.
1const memory = new Memory({
2 options: {
3 titleModel: openai("gpt-4o-mini"), // Override the default model
4 },
5});This gives you more control over memory operations and can help optimize costs by using different models for different memory tasks. #5472
That's it for this week! As always, we'd love to hear your feedback and see what you're building with Mastra.
Happy building! And remember: you can just deploy things. 🚀