Custom agents
Build specialised agents for specific clients or deliverable types.
Custom agents
Two system agents come built-in. Custom agents let you bottle up the way you write for a specific client or deliverable type.
The 30-second version
Each workspace has two system agents: Planner (writes the task plan) and Composer (writes the final deliverable). On top of those, you can build any number of custom agents — each with its own role description, system prompt, model, tools, and Style Memory. Custom agents are how you make Aurora Coffee deliverables sound different from Northwind Real Estate ones.
Why this matters
The reason ChatGPT outputs feel generic is that the model has no persistent memory of how you write. AtelyaOS solves that with custom agents that live in your workspace and carry a Style Memory profile (samples, tone descriptors, banned phrases) the Composer reads on every run.
Custom agents are also where you encode policy. A "B2B SaaS Proposal Drafter" agent can have a system prompt that bans the word "delight" and requires a Pricing & Scope section. A "DTC Caption Writer" agent can be tuned for shorter, punchier output. The Planner picks among the agents available in the workroom when it builds the task plan.
How it works
System agents
Two agents are auto-provisioned per workspace and never appear in the Custom Agents list:
- Planner —
bot_type = 'planner', role: produce a workroom task plan. - Composer —
bot_type = 'composer', role: produce the final deliverable.
You don't manage these directly. They use the workspace's default model and tools.
Building a custom agent
From Agents → New custom agent you provide:
- Identity — name, icon, colour, role description.
- System prompt — the actual instructions the LLM gets before each run.
- Model + temperature — which LLM provider/model to use, plus generation knobs.
- Tools — which integrations or workspace tools the agent may call (
tool_grants). - Context grants — what data the agent is allowed to read.
- Style — uploaded samples + extracted Style Memory profile (see style memory).
[SCREENSHOT: agent builder, Style tab open]
Each agent is versioned (bot_versions table). Editing a custom agent creates a new version; you can view the changelog in the agent detail page.
Agent quotas per plan
Custom agents share a per-workspace limit, controlled by your plan:
| Plan | Custom agents (max_bots) |
|---|---|
| Free Trial | 2 |
| Solo | 2 |
| Starter | 3 |
| Pro | Unlimited |
| Growth | Unlimited |
| Agency | Unlimited |
These caps are enforced at agent-creation time. See pricing for current rates and the full feature matrix.
Picking an agent for a workroom
When you create a workroom, the available agents in your workspace are listed. The Planner uses agents already added to the workroom; the Composer picks the first non-system custom agent in the room as its style source. To make a workroom use a specific agent's voice, add only that agent to the room.
Common pitfalls
- One agent for everything. A single "Generic Writer" agent that handles every client defeats the point. Build at least one agent per major client or deliverable type.
- Long system prompts that contradict Style Memory. Style Memory and the system prompt both feed the LLM. Keep the system prompt about role and policy and let Style Memory hold the voice.
- Forgetting to add the custom agent to the workroom. The Composer falls back to the system default if no custom agent is in the workroom. Add the agent on workroom creation.
- Hitting the agent cap at Solo or Starter. If you serve more than two or three clients with distinct voices, plan to be on Pro or higher.
What's next
- Style Memory — how an agent learns your voice.
- Custom agents per client — the recommended pattern.
- Tool integrations — wiring agents to Notion, Drive, Slack.