Is it possible to use agent + caching ?

[TheUncharted]

Hi
I'm using the Aws Bedrock Provider and the caching is working well with generateText in the example.
But I saw no option for caching for the agent : does this option exists ?

[ahuaiqom]

Hi @TheUncharted 👋

There isn’t a first-class “agent cache” flag today. The Agent is just orchestrating model/tool calls, so you add caching at your app layer—exactly like generateText—and/or cache tool results. (Bedrock itself doesn’t provide response caching.)

**Where to cache

1. Final agent response (safe & simple)**

import { sha256 } from "oslo/crypto"; // or your favorite hasher
import { kv } from "@vercel/kv";      // or Redis/Upstash/etc.

const key = await sha256(JSON.stringify({
  model: bedrock.modelId,
  system,
  tools: Object.keys(toolset),
  messages,            // user + history
  agentVersion: "v1",  // bump when prompts/tools change
}));

let out = await kv.get<string>(key);
if (!out) {
  const result = await agent.run({ messages }); // your Agent call
  out = result.text;                            // or structured output
  await kv.set(key, out, { ex: 3600 });         // TTL 1h
}
return out;

2. Tool-level caching (often biggest win)

const SearchTool = tool({
  run: async (q) => {
    const k = `tool:search:${hash(q)}`;
    const hit = await kv.get(k);
    if (hit) return hit;
    const data = await realSearch(q);           // API call
    await kv.set(k, data, { ex: 86400 });
    return data;
  },
});

3. Streamed agents
Buffer the final message (or structured output) and write it to cache at stream end; serve cached on future identical keys.

Tips

• Key carefully: include model id, system prompt, toolset/version, user input, and any toggles (temperature, maxTokens).

• TTL & scope: only cache responses that aren’t user-private or stateful; separate per-tenant keys.

• Miss tolerance: if a tool mutates state, don’t cache that turn’s final answer.

TL;DR: No built-in agent cache switch; wrap the agent call and (especially) your tools with your store (KV/Redis). That gives you the same wins you saw with generateText, now across the whole agent flow.

[github-actions[bot]]

This discussion was automatically locked because it has not been updated in over 30 days. If you still have questions about this topic, please ask us at community.vercel.com/ai-sdk