PrismerCloud

Prismer Cloud presents itself as an open-source harness for long-running agents, built on the idea that such agents fail without infrastructure for context, error recovery and persistence. The README lists six parts: Context (web content compressed for LLM windows), Memory (working and episodic, persisting across sessions), Evolution (agents learn from each other's outcomes), Tasks (scheduling, retry, cron, exponential backoff), Messaging (agent coordination) and Security. It argues that without these, agents refetch the same URL, forget what they learned and retry blindly.

The most distinctive piece is Evolution. In the README's example, an agent that hits `ETIMEDOUT` calls `runtime.suggest(...)` and gets back a strategy such as exponential backoff with jitter plus a confidence score, then reports the outcome with `runtime.learned(...)` so the shared network improves. It says the selection uses Thompson Sampling that runs in-process with no network call for cached genes. A live Evolution Map is published on the project site.

You can use it through the `@prismer/sdk` npm package and a `prismer` CLI (for example `prismer context load <url>` and `prismer evolve analyze`), an MCP server that the README says needs no API key to explore, or a Claude Code plugin installed with `claude plugin add prismer`. The SDK and CLI need an API key from prismer.cloud, so part of the system is a hosted service. The licence is MIT and the last push was 2026-10-11. The README does not detail what outcome data is shared with the network, so check that before use. This review is based on the README and repository data; we have not run it ourselves.