Working with the Orchestrator
Coordinate allowed helpers and review the results in one conversation.
In this topic
- 01Describe the outcome→
- 02Review the plan or permission→
- 03An allowed agent takes the task→
- 04Read the result
The diagram describes the intended workflow, not a live run.
The built-in Mellow agent is the coordinator for your local setup. Ask it to explain configuration, propose a change, or assign a bounded job to a specialist. Its role differs from a custom agent that directly edits files or runs software.
Start with the current state
A useful configuration request names the intended outcome:
Create an agent for reviewing product documentation. It should read the selected project material, identify unsupported claims, and return an editable draft. Show me the proposed capabilities before changing the setup.
Mellow can inspect configuration and formulate a desired-state change. Review the planned difference, then inspect the apply result. A plan that looks correct and a change that was successfully applied are separate stages.
Configuration tools
| Tool | Role |
|---|---|
mellow_inspect | Read current configuration and schema-backed state |
mellow_help | Retrieve the help material bundled with the app |
mellow_config | Plan and apply supported configuration changes through review |
These tools give the coordinator a concrete source for answers about your setup. Asking “what is configured?” is different from asking a general model question about what an AI assistant could theoretically do.
Delegate a task
The coordinator can use an allowed specialist through spawn_agent. Describe the input, expected output, and boundaries. For example:
Have my Documentation agent compare the draft with the source notes. Ask it to return a discrepancy table and an updated draft, without publishing anything.
A delegate runs with its own agent configuration. Its model, tools, and folder may differ from the coordinator's. Results and shared artifacts return to the parent conversation. Independent tasks can be dispatched together where the configured limits permit.
Working Folder
The coordinator can read and search a selected working folder. It does not gain the custom agent's full file-writing or execution surface. A specialist without its own folder can use the inherited job location; a specialist with a configured folder may use that instead.
Inspect the actual destination before asking for changes. Reading a folder in the coordinator and delegating a write are two different operations with different actors.
Choose allowed specialists
Open the coordinator's Subagents settings to review its allowed list. New custom agents can be seeded into this list through supported creation paths. If you remove a specialist, that choice is preserved rather than treated as a request to refill the list.
For shared agents, inspect the displayed workspace identity. A remote target can be unavailable even while the local coordinator is ready. Do not infer cloud access from the presence of a similarly named local agent.
Customize the coordinator
You can change its display name and add instructions. Those changes do not turn it into a general execution agent. Its protected role continues to govern its available tools.
The model control determines which model answers in the current conversation. Delegates normally use their own configuration unless a supported override is selected. Keep that distinction in mind when diagnosing a specialist's different style or performance.
Repeatable configuration
For a version-controlled setup, the configuration CLI supports exporting, planning, and applying desired state:
mellow config export > mellow-config.yaml
mellow config plan mellow-config.yaml
mellow config apply mellow-config.yaml
Review exported material before sharing it, and manage credentials through their intended secure mechanisms. A configuration file does not provision missing models, external account permissions, or local folders by itself.
Troubleshooting
| Request fails because… | Try this |
|---|---|
| It will not edit a file directly | Delegate to a custom agent with the appropriate folder and tools |
| The intended specialist is absent | Inspect Allowed subagents and the specialist's availability |
| A proposed change did not take effect | Read the apply result rather than the plan alone |
| An external client cannot target the coordinator | Use a permitted saved custom agent; the built-in coordinator is in-app only |
| A delegate used a different location | Compare target-agent folder and inherited working context |
See Subagents for execution budgets and Configuration for desired-state schemas.
Continue exploring · Build your agent teamDelegating to helpers →Understand delegation, parallel work, resource limits and how results return.