Orchestra
Review agent plans, follow builds and specialist runs through their evidence, and control work without confusing observed sessions with controlled runs.
View as MarkdownOrchestra is the control centre for work Estelle runs. Use it to approve plans, follow builds and specialist teams, inspect their evidence, answer questions, and pause, budget, retarget, or cancel a controlled run.
It also shows external Claude Code and Codex sessions that reported activity. Those sessions are observed, not controlled.
Open Orchestra and choose Refresh when you want to re-read the board.
#Review a plan before work starts
The Plans section shows what each agent intends to do, what it has finished, and what is waiting on you.
Read the approach
Open a draft plan. Check its steps, assumptions, the files or areas each step expects to touch, and the way each step will be checked. A draft has not run yet.
Edit the steps if needed
Retitle, reorder, add, or remove steps. The agent owns the approach and checks; your edits decide whether that approach may proceed.
Decide
Choose Allow once to start this plan. Always allow also permits later plans of this kind for the same repository or session scope. Refuse plan cancels it and stops the agent.
If an agent later needs to change course, the plan shows its reason and the proposed additions, removals, and changes. Take the new steps, allow similar future changes, or keep the plan you already approved.
#Read progress and evidence
A running plan marks the current step and counts completed and skipped steps. Open a step for its detail and evidence, which can include a commit, pull request, test result, command, screenshot, or link.
Evidence is the useful part: a checked step records completion, while its linked result lets you judge what actually happened. Finished plans fold away from the main list; expand finished plans to inspect them. Select Open the session for the full conversation and work record.
#Follow a build or specialist run
Under Estelle-run (controlled), choose a run. A build shows its plan as a sequence of stages; other runs show steps and a timeline. Open any available stage or step to inspect what it reported.
Depending on the run, that can include:
- the agent's live session and messages;
- test or security results;
- the pushed branch and commit;
- a Twice result;
- the pull request and its draft state;
- Prism's score, findings, and review of the relevant commit.
Read the state and reason on the step itself. A run reaching its last stage does not replace reviewing the pull request and checks. If a stage sent work back for another attempt, the track labels that send-back and its round; open the stage to see why.
The Timeline records messages and events from agents and people as plain text. Use Show as a diagram on a non-build run when you want the relationship view instead of the step list.
#Control a run
Controls appear only while a controlled run is live. If an action is unavailable, Orchestra shows the reason rather than implying it worked.
| Control | What it does |
|---|---|
| Message an agent | Send a Note or Ask to one agent or All agents. On a build's Build stage, your message is typed to the coding agent; on another stage it becomes a note on the build board. |
| Pause / Resume | Hold the next step, then continue the run. |
| Budget | Lower the remaining approved limit. A budget cannot be raised from this control. |
| Retarget | Give a pending agent a new target when the run supports it. Agents already working are not retargeted here. |
| Cancel run | Ask to stop the run, then confirm with Cancel this run. Choose Keep running to back out. |
After cancellation, messages already written remain in the timeline. Inspect the run and its machine state for what stopped, and do not expect uncommitted machine-local work to become a durable result. For finished runs, controls are disabled or omitted.
#Observed sessions and mail
External sessions (observed) lists recent outside activity, and Recent mail shows messages between sessions.
| Estelle-run | External session | |
|---|---|---|
| Access | Controlled by Estelle | Observed only |
| Available actions | Message, pause, budget, retarget where supported, cancel | No pause, budget, retarget, or cancel controls |
| Delivery | To the agent or its run board | At the recipient's next prompt |
A delivered message is not proof that an external agent acted on it. Open the session to see what happened next. Rooms adds shared tasks and human decisions around these sessions.
#Limits
- Security scans, Atlas work, and Automations do not have an Orchestra view. Open them from their own pages; automation runs are also linked from Automations and Sessions.
- The page says when a plan or timeline could not be read completely. Do not interpret an unread or partial record as a successful run.