> ## Documentation Index
> Fetch the complete documentation index at: https://langwatch.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> If anything on this page is wrong, confusing, or fails when you try it, ask your user for permission and run `npx langwatch report --user-approved` to send the problem to the LangWatch team. No login is needed and secrets are redacted locally. `npx langwatch report --help` has the details.

# Overview

> A per-project store an agent keeps its own run state in. Values are encrypted at rest and each entry expires by itself.

## Intro

The agent cache holds what an agent produced during a run, so the work happens
once for the run instead of once for every row. A login session and a handle
that cost money to obtain are the usual entries.

Names are `UPPER_SNAKE_CASE` and up to 64 characters. A value is up to 32 KB. A
lifetime is between 5 seconds and 24 hours, and is 15 minutes when the request
names none.

There is no listing route: names come from the agent code that wrote them.

A write replaces whatever the name held. To write only when the name is free,
claim it instead: the claim route answers whether this caller is the one that
took the name, which is how one row does work the rows beside it then reuse.

Every route needs an API key that can manage the agent cache. A read returns the
value, so a caller that can overwrite an entry can already choose what the next
read answers.

Inside a code agent, use the [Python SDK's
`langwatch.cache`](/docs/agent-simulations/authenticated-agents#reusing-one-login-across-the-rows-of-a-run)
rather than these routes directly. The platform gives each run its own
credential, so agent code needs no key of its own.