Control plane for coding agents
Publish a focused agent. Delegate from Linear or GitHub. Follow isolated runs from task to human review.
Make waiting states
easier to spot.
Frontend maintainer
Frontend maintainer v3
Linear task attached to a dedicated session.
Waiting states clarified. Changes ready for human review.
From delegation to review
A control plane is the place your team defines an agent, gives it repository access, sends it work, and follows what happens next. Terrarium connects those decisions to a run instead of leaving the result in a separate conversation.
For engineers, engineering leaders, and platform teams delegating scoped repository work. Keep Linear and GitHub as the source of the task; use Terrarium to configure execution and inspect the handoff.
Instructions, model, capabilities, and allowed repositories.
A request with an outcome, context, and constraints.
Its own session and sandbox, with visible state and timeline.
Inspect available evidence and any linked output.
01 / Define the agent
Give an agent a specific job and durable operating instructions. Choose a supported model, inspect its available capabilities, and attach only the repositories it should work with.
Publishing creates an immutable version. New runs use a published version, and the version used stays attached to each run—even when you later edit the draft.
About model selectionFocused interface changes in example/workspace
System instructions · Example excerpt
Choose from the supported language models currently listed in the editor. Capabilities come from the platform catalog and reviewed organization connections, not arbitrary uploaded code.
02 / Send the task
Start directly in Terrarium, use a Linear Agent Session, or mention the installed app in a GitHub conversation. Provider-originated work stays connected to its source.
Choose a published Coding Agent and an allowed repository. Confirm the branch and access, then describe the change and expected validation.
Alternative launch · EXAMPLE-24
Frontend maintainer v3 · example/workspace
Make waiting states easier to spot. Check the narrow layout and keyboard focus.
An alternative manual launch—not a second request in the Linear run below.
Delegate an issue to the installed app, or mention it in a comment to start a supported Agent Session. Follow the task in its issue context.
Example issue · EXAMPLE-24
Make waiting states easier to spot.
Delegated to Frontend maintainer · This is the source used in the story below.
Mention the installed app in an issue or pull request conversation. Give it a bounded request tied to that discussion.
Alternative entry · Example conversation
Please check the waiting-state treatment in this change, including keyboard focus.
An alternative launch path—not another event in EXAMPLE-24.
Provider setup comes first. GitHub repository access and Linear workspace authorization must be in place. New native GitHub and Linear app setup may need an operator; it is not universally self-service. These are supported intake paths, not support for every provider event. See prerequisites ↓
03 / Follow one connected run
EXAMPLE-24 follows Frontend maintainer v3 in example/workspace. This fictional sequence ends with the possible pull request shown above. It illustrates the record—not a promise that a run will finish or produce code.
EXAMPLE-24 / Selected context
“Make waiting states easier to spot.”
Manual and Linear attempts capture redacted prompt and context snapshots. Inspect what is available; a snapshot is not promised for every intake path.
The run records the source task and published agent version.
A dedicated session and Vercel Sandbox prepare the selected repository and branch with configured access. Conversation state and working files are separate from other tasks.
In this illustration, output is recorded as pull request #42, “Clarify waiting states.” A run can expose commit or pull request links when it produces them.
Read the output, inspect the diff, verify checks and behavior, and decide whether to accept the change. A completed run is not an automatic merge.
Other possible states
These are alternatives to the completed example above, not later events in its timeline.
A run may pause for clarification or approval. Inspect the request and provide the missing context. Active or waiting runs with an executor session can receive follow-up input.
Review the recorded state and available timeline to understand where work stopped. Completed, failed, and canceled runs cannot receive follow-up input. Do not assume an output exists.
04 / Stay in control
Isolation, repository settings, and provider credential separation each have a specific job. They do not replace clear task instructions or your team's code review.
Current limit · Linear updates
Agent-initiated Linear updates through the platform connection are currently unavailable.
This limit concerns updates, not Linear task intake. Task instructions such as “keep changes focused” are not per-file enforcement.
Practical starting points
Hypothetical engineering tasks you might delegate—not packaged features, customer stories, or guaranteed outcomes.
Requested scope guides the task. Repository read/write settings are not per-file restrictions.
Getting started
You need a Terrarium account and organization, access to the intended GitHub repository, and the appropriate permissions to configure connections, repositories, and agents.
Linear intake also needs an authorized workspace and installed app. Confirm provider availability with your organization's operator before relying on a channel.
Open TerrariumOpens the existing Terrarium sign-in.
Use your existing organization, or create an organization through workspace setup. Ask an owner or admin if you need access or configuration permissions.
In Connections, confirm organization GitHub access. In Repositories, import a repository visible to that installation; set its base branch, read/write access, and any setup and validation commands.
Set its model and instructions, review available tools, and attach allowed repositories. Save the draft, then publish a version. Per-agent provider setup lives in Channels; GitHub and Linear setup may require an operator.
In Runs, select the published agent and allowed repository, confirm the branch and access, and enter task instructions. Follow status and timeline, respond when input is needed, and review any linked output.
Before you delegate
The agent editor lists supported language models from AI Gateway. Select from that current catalog rather than assuming every model or provider is available. The Claude Sonnet label in this page is an illustrative selection, not a fixed model list. Your chosen model is part of the published agent version.
Do not assume universal self-service setup. Organization connections and per-agent native Channels are separate. GitHub needs installation access to the intended repository; Linear needs workspace authorization and an installed app. New native GitHub and Linear app setup can require an operator, and Linear installation is not available to every customer workspace. Check the setup state with your organization's operator.
The organization, published agent's allowed repositories, branch settings, and repository read/write setting constrain its target and write actions. A manual run selects an allowed repository and can specify a starting branch override. Instructions can ask for a narrow change, but they are not per-file permissions.
No. A run can expose linked commits or pull requests when it produces them. Inspect available artifacts and references in its record; some sections may be empty until relevant data is emitted. A completed status does not prove that checks passed or that a change is ready to merge.
Waiting means the executor is paused at a boundary and may need context or approval. Inspect the request before responding. Failed means execution stopped before completion: review the available timeline and error context, correct the cause, and decide whether to start a new task. No commit, pull request, or successful validation is guaranteed.
No. Follow-up input is available for active or waiting runs with an executor session. Completed, failed, and canceled runs cannot receive it. If you need more work after a terminal state, start a new task with the relevant context.
Define the task and access, answer clarification requests, and verify the output. Check the code, tests, behavior, and remaining risks before accepting a change. Agent-initiated Linear updates through the platform connection are currently unavailable; Linear Agent Sessions remain a task entry path. Your team owns merge decisions; Terrarium does not automatically merge, deploy, or release code.
Scoped work. Visible handoff.