is it possible to run claude code directly on a agent folder?

[lfnovo]

Sometimes it makes more sense to be able to talk to an agent directly through Claude Code to do some work that is not ready for an issue or to develop memories more profoundly. How can one do that?

[Jeanlaris]

Hello @lfnovo,

Executing agentic workflows directly via code to handle unpublished work and build deeper memory persistence requires decoupling the execution loop from the public repository lifecycle.

Here is how you can architect this approach:

1. Ephemeral Development Branches: Route your agentic sessions to local staging branches or isolated workspace directories. This allows the model to iterate on raw, uncommitted code without polluting main production lines.
2. Externalized Memory State (RAG & Vector Stores): To develop deeper memories beyond standard context windows, attach a local vector database (like Chroma or FAISS) or persistent JSON/SQLite memory layers. The agent queries and updates these states programmatically during execution.
3. Programmatic Agent Hooks: Use direct API calls or local client SDKs (such as Anthropic's SDK) combined with custom system prompts that define read/write boundaries for your local file system, ensuring local experimental code stays strictly contained until ready.

How are you currently managing the memory persistence layer in your local agent loops? Let's exchange architectures via our repositories!

[lab1207]

Yes — an agent's folder is a normal checkout, so run any supported CLI directly inside it (claude, codex, cursor all have adapters). Paperclip only tracks work that flows through issues and sessions, so local unpublished exploration stays yours and never touches the board.

Two rules to avoid trouble:

1. One writer per task. The task model is single-assignee with atomic checkout. Don't have the Paperclip agent and your local CLI working the same checked-out issue at the same time — they will step on each other's changes. Do the local work first, commit it, then hand the issue to the agent (or vice versa).
2. Bring it back through git. Commit and push the local work, then point the agent at the branch. It continues from your commits instead of redoing them.

For the memory side: anything the CLI writes into the repo (notes, docs, memory files) persists like any other file, so deeper context you build locally carries over whenever the agent next checks out that folder.