Nothing reaches an agent, and nothing leaves it, except through a gate you configured. Four mechanisms make that true.
An agent in Cast is a directory on your disk. Open it in Finder. The instructions, skills, and memory are plain files you can read. blueprint/ is mounted read-only, memory/ is writable. The agent updates its memory and working files as it works, but never edits the prompt that tells it who it is, because the harness doesn't let it.
Read more →Every agent runs in its own container. Scoped filesystem: only the folders you mounted are visible. Scoped network: only the hosts on its allowlist are reachable. Tools are declared per channel, so the same agent can have a full surface in your main conversation and a narrow one in an automated lane. A clever model can't argue past it.
Read more →A channel in Cast isn't just a way to reach the agent. It's a room with its own rules. Default for real conversations. Email triages incoming mail with a narrowed tool surface. Scheduled runs overnight maintenance with no user interaction. Peer is how this agent talks to another. Each room has its own tools, lifecycle, and memory conventions.
Read more →Agents can discover each other, but reaching is gated. An edge is per-channel and directional, and it becomes real only when the receiver's owner approves it. The sender needs reach, the receiver needs to accept. Until both are there, nothing crosses. The cook-agent can ask the calendar what's on tonight, but it has no path to the journal.
Read more →