Blog
Notes from the Tuist team.
Release notes, OTP patterns, and lessons we learn while building production AI agents on the BEAM.
-
Microsandbox support in Condukt
Condukt can now run agent tool calls inside a local microVM through Microsandbox, with the same tools and the same agent definition.
-
The network is the open side of the sandbox
Sandboxing the filesystem and processes is the easy part. The network is the surface most stacks leave open, and the one that matters most as agents start touching real systems. Here is how we are closing it in Condukt, and the missing piece we are watching for.
-
Coding agents as runtimes
Codex and Claude are already tuned coding harnesses. Condukt should meet them where they are and orchestrate them as runtimes.
-
Tool calls in a pod you already run
We just shipped a Kubernetes sandbox for Condukt. One pod per session, the same agent definition, the same tools.
-
Where an agent's tools should run
We just landed sandboxes in Condukt. Same agent, swap where the tool calls actually execute. Here is the thinking behind it.
-
Secrets belong in the session
Agents need to act against real systems. We added session secrets to Condukt so credentials become part of the execution boundary, not part of the conversation.