🤔 Problem

A recent proof of concept demonstrated the usefulness of Renku for agentic coding, because it can enable secure, isolated environments and reduce the attack surface by minimising exposure to secrets, e.g. git credentials can isolated from the running environment and we can run dedicated sessions for different projects.

We need to enable use of agents in Renku in an easy and convenient way.

Based on ‣.

🍴 Appetite

6 weeks

🎯 Solution

• We need buildpacks that will install specific agents when environments are build from code.
• UX/UI flow for using agents in Renku sessions.
• Security precautions and sandboxing.
  • see e.g. https://github.com/carderne/pi-sandbox

🚞 User stories / journeys

• As a Renku user I want to have a convenient way of adding coding agents to my Renku session.
• As a Renku user I want to have a convenient way of authenticating with the coding agent I use. Ideally I would not have to repeat the authentication for every project or every time I start a session.
• As a Renku user I want to have some sandboxing or security guarantees when I use coding agents in a Renku session. For example:
  • I dont want the agent to delete code on remote repositories
  • I dont want the agent to exfiltrate or read secrets
  • I dont want the agent to delete data from data connectors mounted in Rread/Write mode.
• As a Renku user I want to have access to a variety of coding agents:
  • Claude Code
  • Codex
  • Pi
  • Opencode
  • Models hosted by my institute or company for internal use. (This also includes the LLMs we are hosting at the SDSC or SwissAI)
• I want the context of the agent to remain and be preserved when I pause and resume a session.

🐰 Rabbit Holes

• Many different models and providers. Not sure if there is an underlying standard that is common.