Coding agents

Use coding agents to create and evolve fully managed Redgold applications. Agents propose changes to schemas, pipelines, transforms and tests; Redgold operates the application data, runtime and deployment. Agents also support authorized external repositories and their existing delivery systems.

A coding-agent session combines a conversation with an isolated repository workspace and an allowed set of tools. Use it for work that benefits from inspecting files, making coordinated edits, and running verification in the repository environment.

Give the agent a bounded task

A useful task names the desired outcome, the relevant area of the repository, and any constraints on scope or verification. The agent can inspect local project instructions and existing patterns before changing files.

For broad work, start with an audit or plan in the same conversation. This makes assumptions and affected components visible before implementation begins.

Review the run

Messages and tool activity remain associated with the session. Review the changed files and verification output as evidence. Persistence helps with continuity and audit; it does not guarantee that an LLM replay produces identical output.

Delegate independent work

An agent can delegate bounded investigations that can run independently, such as reviewing separate crates or checking documentation and API generation in parallel. The parent conversation remains responsible for reconciling results and verifying the combined change.

Organize ongoing application work

For work spanning multiple tasks or sessions, use Agent operations. The guides cover workstreams and issues, agent lifecycle, and supervision and recovery. A worked application change follows one objective through a blocked decision, agent work and managed delivery.

Sensitive actions

Authentication, permissions, and approval policy determine which external or destructive actions a session may take. Keep credentials outside prompts and repository files. Human review remains appropriate for production changes, access grants, and irreversible operations.

Start a session from the quickstart, then use AI and agents for the surrounding platform behavior.

Watch the agent over SSH

Each session runs in its own managed pod with the agent process inside a tmux session, and SSH access to that pod is available on enabled accounts. When it is enabled for you, your account's agent view shows a connect string of the form ssh <user>@ssh.redgold.ai, authorized against the public keys registered on the account. Inside the pod, tmux ls lists the live agent sessions and tmux attach -t <name> puts you at the agent's actual terminal — you can watch it work, take over the keyboard, and detach again without ending the run. The workspace checkout and shell state are right there too, so anything the transcript leaves ambiguous you can inspect directly. The web and terminal clients render the same pane output and session state, so what you see over SSH matches what the platform surfaces elsewhere.

How this maps to the platform shape

The agent plane is a vertical built on the same base the rest of the platform runs on: agent sessions, tool calls, and model responses are signed rows with content-addressed keys, and the surfaces that query them are dataflow pipelines. Building a vertical of your own follows the same path — Build a vertical walks it in order, and Managed application artifacts explains the files reviewed during the managed preview.