How to Fix 'state not updating in production' in AutoGen (TypeScript)

If your AutoGen TypeScript agent works locally but stops updating state in production, you’re usually dealing with a lifecycle or persistence bug, not a model bug. The common pattern is: messages arrive, the run completes, but your stored conversation state never reflects the latest turn.

In AutoGen TS, this usually shows up around AssistantAgent, UserProxyAgent, Memory, or your own wrapper around run()/onMessage(). The error text is often indirect: state not updating, stale ThreadState, missing memory writes, or a production-only gap where the agent responds but your app-level state stays frozen.

The Most Common Cause

The #1 cause is mutating local state inside a callback or request handler, then expecting that mutation to survive across async boundaries or serverless invocations. In production, especially under Node serverless, workers, or React/Next.js API routes, that in-memory object gets recreated or isolated.

Here’s the broken pattern:

import { AssistantAgent } from "@autogen/core";

let conversationState = {
  messages: [] as Array<{ role: string; content: string }>
};

const agent = new AssistantAgent({
  name: "support-agent",
  systemMessage: "You are a support agent.",
});

export async function handleChat(input: string) {
  const result = await agent.run({
    messages: [
      ...conversationState.messages,
      { role: "user", content: input },
    ],
  });

  // Broken: assumes this mutation will persist reliably in prod
  conversationState.messages.push({ role: "user", content: input });
  conversationState.messages.push({ role: "assistant", content: result.output });

  return result.output;
}

And the fixed pattern:

import { AssistantAgent } from "@autogen/core";

type ChatMessage = { role: "user" | "assistant"; content: string };

const agent = new AssistantAgent({
  name: "support-agent",
  systemMessage: "You are a support agent.",
});

async function loadConversation(threadId: string): Promise<ChatMessage[]> {
  // Replace with Redis / DB / durable storage
  return [];
}

async function saveConversation(threadId: string, messages: ChatMessage[]) {
  // Replace with Redis / DB / durable storage
}

export async function handleChat(threadId: string, input: string) {
  const previousMessages = await loadConversation(threadId);

  const result = await agent.run({
    messages: [...previousMessages, { role: "user", content: input }],
  });

  const nextMessages = [
    ...previousMessages,
    { role: "user", content: input },
    { role: "assistant", content: result.output },
  ];

  await saveConversation(threadId, nextMessages);

  return result.output;
}

The key difference is simple:

• •Broken code treats process memory as persistence
• •Fixed code treats storage as the source of truth

If you’re seeing something like AssistantAgent.run() returning valid output while your UI still shows old state, this is almost always the reason.

Other Possible Causes

Missing await on persistence

// Broken
saveConversation(threadId, nextMessages);
return result.output;

// Fixed
await saveConversation(threadId, nextMessages);
return result.output;

If your logs show 200 OK before your DB write completes, you’ve found it.

Multiple worker instances

// Broken in multi-instance deployments
const cache = new Map<string, ChatMessage[]>();

// Fixed
// Redis-backed store keyed by threadId

This bites teams deploying to Vercel, Kubernetes replicas, or any autoscaled Node service. One instance writes the message; another instance handles the next request and sees empty state.

Wrong session key

// Broken
const threadId = req.headers["x-session-id"] as string;

// Fixed
const threadId = String(req.body.threadId ?? req.headers["x-session-id"] ?? "");
if (!threadId) throw new Error("Missing threadId");

A mismatch here produces classic symptoms:

• •user sends message A
• •assistant responds correctly
• •next turn starts from an empty context

Stale closure in callbacks

// Broken
function onNewMessage(msg: ChatMessage) {
  setMessages([...messages, msg]); // messages may be stale
}

// Fixed
function onNewMessage(msg: ChatMessage) {
  setMessages((current) => [...current, msg]);
}

This is common in React apps wrapping AutoGen agents. The agent is fine; your UI state update path is stale.

How to Debug It

1. •

   Log the thread ID and message count at every boundary

   • •Before agent.run()
   • •After agent.run()
   • •Before save
   • •After save

   You want to confirm whether the bug is in AutoGen execution or in your persistence layer.

2. •

   Check whether the same process handles both requests

   • •Add process.pid to logs.
   • •If turn one and turn two hit different PIDs and you’re using memory-only storage, that’s your answer.
3. •

   Inspect the exact data passed into AssistantAgent.run()

   • •Print the array length and last message.
   • •If the prompt contains stale messages before the model call, the bug is upstream of AutoGen.
4. •

   Verify persistence independently

   • •Write a test that calls save/load without AutoGen.
   • •If load returns old data there too, stop debugging agents and fix storage first.

A good debug log looks like this:

console.log({
  pid: process.pid,
  threadId,
  beforeCount,
  afterCount,
});

If beforeCount resets unexpectedly between requests, you’re not dealing with an AutoGen state bug. You’re dealing with deployment architecture.

Prevention

• •Use durable storage for conversation state:
  • •Redis for low-latency session memory
  • •Postgres for auditability and replay
• •Treat AutoGen agents as stateless executors:
  • •Load state in
  • •Run agent
  • •Persist state out
• •Add integration tests for multi-request flows:
  • •First request creates context
  • •Second request must see prior turns

If you’re building anything beyond a toy demo, do not keep conversation history in a module-level variable. That works on localhost and fails exactly when traffic moves to production.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.