AgentNetwork (Experimental) #2744
Conversation
Review or Edit in CodeSandboxOpen the branch in Web Editor • VS Code • Insiders |
|
|
The latest updates on your projects. Learn more about Vercel for Git ↗︎
|
|
📝 Documentation updates detected! You can review documentation updates here |
| // const { memory } = useMemory(networkId!); | ||
| // const navigate = useNavigate(); | ||
| const { messages, isLoading: isMessagesLoading } = useMessages({ | ||
| agentId: networkId!, | ||
| threadId: threadId!, | ||
| memory: false, | ||
| }); | ||
|
|
||
| // const [sidebar, setSidebar] = useState(true); | ||
| // const { | ||
| // threads, | ||
| // isLoading: isThreadsLoading, | ||
| // mutate: refreshThreads, | ||
| // } = useThreads({ resourceid: networkId!, agentId: networkId!, isMemoryEnabled: !!memory?.result }); | ||
|
|
||
| // useEffect(() => { | ||
| // if (memory?.result && !threadId) { | ||
| // // use @lukeed/uuid because we don't need a cryptographically secure uuid (this is a debugging local uuid) | ||
| // // using crypto.randomUUID() on a domain without https (ex a local domain like local.lan:4111) will cause a TypeError | ||
| // navigate(`/networks/${networkId}/chat/${uuid()}`); | ||
| // } | ||
| // }, [memory?.result, threadId]); |
There was a problem hiding this comment.
Do we need this commented code? Should it be uncommented or deleted?
| {/* {sidebar && memory?.result ? ( | ||
| <div className="border-r"> | ||
| <div className="p-4"> | ||
| <h3 className="text-sm font-medium text-mastra-el-5 mb-2">Network Conversations</h3> | ||
| {isThreadsLoading ? ( | ||
| <p className="text-sm text-mastra-el-4">Loading threads...</p> | ||
| ) : threads.length === 0 ? ( | ||
| <p className="text-sm text-mastra-el-4">No conversations yet</p> | ||
| ) : ( | ||
| <div className="space-y-2"> | ||
| {threads.map((thread) => ( | ||
| <div | ||
| key={thread.id} | ||
| className={`p-2 rounded-md cursor-pointer text-sm ${thread.id === threadId ? 'bg-mastra-bg-3 text-mastra-el-5' : 'hover:bg-mastra-bg-2 text-mastra-el-4'}`} | ||
| onClick={() => navigate(`/networks/${networkId}/chat/${thread.id}`)} | ||
| > | ||
| {thread.name || 'Conversation'} | ||
| </div> | ||
| ))} | ||
| </div> | ||
| )} | ||
| </div> | ||
| </div> | ||
| ) : null} */} |
| type AgentTool = ReturnType<typeof createTool>; | ||
|
|
||
| // Create tools from agents with state tracking | ||
| const agentAsTools = this.agents.reduce<Record<string, AgentTool>>( |
There was a problem hiding this comment.
Thought for a later day: during the demo earlier the multi-step agent calling was a little slow and it just occured to me we could give the router agent a parallel agent tool so it could message multiple agents, they could generate at the same time, and it could get the result back. ex input args where each key is the name of an agent and the value is the message to send.
{ agentA: "do x", agentB: "do y" }
Res would match:
{ agentA: "As a large language model...", agentB: "lol" }
There was a problem hiding this comment.
nice thing is this could be a single tool then. The router could choose to call one or more agents with it
|
|
||
| const result = await agent.generate(agentContext, { | ||
| context: this.current_state?.lastResult | ||
| ? [{ role: 'assistant', content: this.current_state?.lastResult }] |
There was a problem hiding this comment.
not important for this PR but we could use memory for this
| getNetworkState: createTool({ | ||
| id: 'getNetworkState', | ||
| description: 'Get the current state of the agent network, including previous agent results and call history', | ||
| inputSchema: z.object({}), | ||
| outputSchema: z.object({ | ||
| state: z.object({ | ||
| callCount: z.number(), | ||
| lastResult: z.any(), | ||
| input: z.string(), | ||
| agentResults: z.record(z.string()), | ||
| agentHistory: z.array( | ||
| z.object({ | ||
| agent: z.string(), | ||
| input: z.string(), | ||
| result: z.string(), | ||
| timestamp: z.string(), | ||
| }), | ||
| ), | ||
| stateData: z.record(z.unknown()), | ||
| }), | ||
| }), |
There was a problem hiding this comment.
This looks like it could use a huge amount of tokens - isn't this covered by the agent tool responses in each agent tool call? Since the tool call args and results will be in context for the router agent
Vibe coded this! It works mediocrely. Will have to make it better