Help center
SW Agent's agents: choose one, or let SW Agent choose
10 agents, each with its own tools, and Auto: SW Agent answers itself, hands a job to the agent whose tools fit, or splits a long request among several and reports what each did.
Describes Sonic Weaver v5.31, released .
What an agent is
An agent is a part of SW Agent that does one kind of work with its own tools: the Writer agent writes, the Designer agent makes pictures, the Audio agent gets your sound ready. Each one is made from the functions, workflows, apps and workspaces SW Agent already has, so it does exactly what those tools do, and every new tool of its kind joins it.
The chat is the controller. It reads your message, then answers it itself, hands it to one agent, or splits it among several. Agents never call each other: SW Agent does every hand-off, and says in the reply's thinking steps, its Ran line and its node graph which agent did what.
Auto or an agent
Auto is the best choice for most requests. SW Agent chooses for each request: it answers itself, hands the job to the agent whose tools fit, or splits a long request among several.
Choose an agent when you want only its kind of work: the Writer agent writes and never builds a project, the Planner agent shows the steps and changes nothing until you press Run.
A chosen agent still lets SW Agent answer your questions. A job outside its tools is offered to the agent that has it, never done by a guess.
For one request only, name the agent at the start ("writer agent: a post about my album") and your choice stays as it was.
- Use the Agent menu beside the message box, in SW Agent's page, the editors' SW Agent tab and the SW Agent panel on every tool page.
- Type "/agent writer" to choose an agent, "/agent auto" to let SW Agent choose, and "/agents" to see them all.
- Or say it: "use the designer agent", "back to auto", or for one request only "writer agent: a post about my album" or "@designer a logo for Night Drive".
- Settings, SW Agent, Agents sets your default. The choice is kept on this device.
The agents
Auto: SW Agent chooses for each request: it answers itself, hands the job to the agent whose tools fit, or splits a long request among several. Open the Agents view in SW Agent to see each agent's rules, tools and workflows, start a workflow, or try one of its requests.
- Visualizer: Finds the right look for your song, builds it as a project, sets up its scene and tunes it. It has 8 functions, 8 workflows and 0 apps. Try "make a neon tunnel visualizer for my synthwave track".
- Editor: Makes the edits you describe in Classic and Studio: controls, layers, keyframes, audio reactions and the project's settings. It has 5 functions, 1 workflow and 0 apps. Try "make the glow brighter and rename layer 2 to chorus".
- Writer: Writes from your words: blog posts in the structure and style you choose, titles, social text, names, documents and a check of your own writing. It has 6 functions, 2 workflows and 12 apps. Try "write a blog post about making my first music video".
- Designer: Makes the pictures around your music: logos, thumbnails and release sets, cover art, stream screens, watermarks, palettes and type. It has 3 functions, 9 workflows and 4 apps. Try "make a logo for my band Night Drive in purple".
- Audio: Gets your sound ready: mastering, stems, beats and sound effects, captions and SRT files, lyrics, rhymes and chords. It has 2 functions, 2 workflows and 8 apps. Try "master this song for spotify and split the stems".
- Researcher: Answers your questions from every source that serves them, leading with the answer, and draws charts and works out the maths. It has 5 functions, 0 workflows and 4 apps. Try "what does the research say about music and sleep".
- Library: Finds copyright-free and openly licensed pictures, sounds, music and books in the free library, and makes things from them. It has 2 functions, 5 workflows and 1 app. Try "find a free rain sound effect".
- Script: Automates the other tools: runs the built-in scripts and yours, writes a script from your words, and puts expressions on controls. It has 4 functions, 0 workflows and 0 apps. Try "write a script that makes my project in every frame shape".
- Export: Gets your work out: renders the open project to a video file, makes PDFs, works out sizes and bitrates for an upload, and gives embed codes. It has 3 functions, 2 workflows and 2 apps. Try "export my project to an mp4".
- Planner: Plans without acting: shows the steps any agent would take for your request, and runs them only when you press Run. Its own planning tools (checklists, storyboards, repurpose plans, script outlines) change nothing, so they answer at once. It has 0 functions, 0 workflows and 5 apps. Try "plan a release for my album".
A long request, split among agents
When you ask for several things at once ("write a blog post about my album and make a logo for my band"), SW Agent plans the steps, hands each to the agent whose tools it uses, and runs them as one job. A part that needs what another made waits for it and is given it (in order); parts that need nothing from each other run side by side, so one that cannot finish never stops the others.
When it ends, SW Agent collects what each agent made, checks each result against what it should be as it comes back, and reports: what each agent did, which finished and which did not and why. A step that did not finish is offered again on its own, with one tap.
If you chose one agent, only its parts run, and SW Agent says which parts it left out and which agent would do them.
Stopping, approvals and limits
An agent's job runs through SW Agent's worker, like every chain and job: Stop takes effect at the next checkpoint, Jobs can pause and resume it, and its record lists every step. In Classic or Studio, a step that would change your open project or start a long render asks you first, in the chat and in Jobs, whichever agent does it.
Each agent keeps to its limits for one job: a number of steps, a time (not counting a pause or waiting for your OK) and a budget of requests to our own server. A step past a limit does not run and says why, and the other agents carry on. SW Agent never spends money: it calls no paid service and no AI model.
The Planner agent changes nothing: it shows the steps any agent would take, naming the agent for each, and runs them only when you press Run.
With other AI assistants
Assistants connected to Sonic Weaver's MCP server can read every agent, with its rules, tools, workflows, approvals and limits, through list_agents. It is read-only: an agent runs in your browser, from SW Agent's chat, and the listing gives the link that opens SW Agent with a request for that agent ready to send.
Still have a question about this?
SW Agent answers from this article and the rest of our docs, and can do it for you with the site's own tools.
Help center
Keep going
Was this page helpful?
Anonymous. Only this page’s address and your yes/no are recorded, and only if you allowed analytics cookies.