Every request you type into the Rixoffice composer is handled by a small, named team rather than one anonymous model. Knowing who is on that team — and what each member is good at — helps you write better requests, understand the results you get back and predict where the work will be strong or weak. The roster is small on purpose: five agents, five clear roles, one shared set of app skills.
Rixoffice ships five agents. Each has a distinct role, and they collaborate through delegation:
| Agent | Role | Strength |
|---|---|---|
| Sisyphus | Orchestrator (default) | Routing, splitting and supervising work |
| Hephaestus | Executor | Carrying out edits and builds with tools |
| Prometheus | Planner | Structuring long or complex jobs before execution |
| Momus | Reviewer | Critiquing output and requesting fixes |
| Oracle | Read-only analyst | Answering questions without touching the file |
You will rarely interact with anyone but Sisyphus directly. The others are reached through delegation, which is exactly how it should feel: a manager assigns the work, you talk to the manager.
Sisyphus, the orchestrator
Sisyphus is the agent you are always talking to. It runs by default in every app. When your request is small — fix this sentence, explain this formula — Sisyphus handles it directly, choosing the right tools in one pass. When your request is big, Sisyphus becomes a manager: it breaks the job into subtasks, hands each one to the right specialist with the agent_delegate tool, watches the budgets and assembles the final result into a coherent answer.
Because Sisyphus is the default, you never have to choose an agent to get started. The routing is invisible when it does not matter and visible in the composer when it does, so you can always see who was assigned what.
Hephaestus, the executor
Hephaestus does the hands-on work. Give it a plan or a direct instruction and it calls the app's skill tools to make real changes: inserting blocks in a document, proposing workbook operations, building slides from a script, marking up PDF pages. Hephaestus is the agent behind every edit that actually lands in your file, which is why its output is always visible in real time — you watch the executor work rather than waiting for a finished blob.
Executor quality depends on clarity of instructions, and that is exactly why the planner and reviewer agents exist around it.
Prometheus, the planner
Some jobs should not start with editing. Ask for a twenty-slide deck or a restructuring of a long report and Prometheus first produces a plan — an outline, a page structure, a sequence of steps — that the executor can then carry out. Planning first is what keeps big generations coherent instead of rambling: the executor always has a skeleton to fill rather than improvising page by page.
You get to read and adjust the plan before anything is written into your file. Treat the plan as a contract: if the outline looks right, the execution usually lands right.
Momus, the reviewer
Momus exists because first drafts are drafts. After work is executed, Momus can review the result against the request and ask for fixes: a layout audit on generated slides, a consistency pass on a rewritten section, a check that a table really matches the underlying data. This review loop is built into flows like AI deck generation, and it is why the slide QC agent can catch and repair problems before you ever see the deck.
The reviewer closes the loop that pure generate-and-pray AI leaves open. It is also the agent most likely to save you from shipping a subtle mistake.
Oracle, the read-only analyst
Oracle answers questions. It can read your document, workbook, deck or PDF and tell you what is inside: summaries, data checks, formula explanations, comparisons between sections. The defining property is in the name — read-only. Oracle has no tools that modify your file, so asking it for an opinion is completely safe.
Use Oracle whenever you want understanding without change: what does this document claim, where do these two tables disagree, is the logic of this argument sound. It is also the natural agent for reviewing someone else's work before you comment on it.
How delegation flows in practice
- You type: review the annual report and fix the weak sections.
- Sisyphus reads the document and splits the request into subtasks.
- Prometheus is delegated the outline of what needs fixing.
- Hephaestus applies the approved fixes block by block.
- Momus reviews the result and requests a fix for one weak table.
- Sisyphus collects everything and reports exactly what changed.
Every delegation in that chain carries a budget, so the whole job stops cleanly at its limits instead of wandering. Two habits make your requests work better with this pipeline: say what the outcome should be, not only the action, and name the target when you can — a heading, a sheet, a slide — so the orchestrator routes the work precisely. Clear targets mean fewer delegated steps and faster results.
What to explore next
- Agent skills — the exact toolsets each app gives these agents.
- AI agents overview — the loop and delegation model behind the roster.
- Slide QC agent — Momus at work in RixSlides.