AtelyaOS Docs
Customization

One custom agent per client

The recommended pattern for agencies serving more than two clients.

One custom agent per client

Different clients want different voices. Build one agent per client and stop fighting your style guide.

The 30-second version

Build one custom agent per major client. Each agent has its own system prompt (role, policy) and its own Style Memory profile (samples, tone, banned phrases). When a workroom belongs to that client, add their agent to the workroom — the Composer uses their voice, not the workspace default. Pro and above let you keep an unlimited number of agents.

Why this matters

A single "agency writer" agent is a compromise. A B2B SaaS proposal voice and a DTC fashion caption voice cannot coexist in one Style Memory profile without averaging into mush. Splitting them is the cleanest solution: each client gets a stable, inspectable voice that doesn't drift between sessions.

This is also where you encode client-specific policy. Aurora Coffee always wants their tagline in the footer; Northwind Real Estate's compliance team bans the word "guaranteed". Those rules live in the agent, not in your head.

How it works

When to make a per-client agent

Create a client-specific agent when any of the following are true:

  • The client's voice is materially different from your house default.
  • The client has do/don't rules that need to be enforced on every deliverable.
  • You'll deliver more than two pieces of work to them per quarter.

For one-off work or a quick recap, the workspace default agent is usually fine.

Setting up the agent

  1. Go to Agents → New custom agent.

  2. Identity: Aurora Coffee — Proposal Drafter. Pick an icon and colour.

  3. Role description: one paragraph stating what kind of work this agent produces.

  4. System prompt: include client-specific policy, e.g.:

    You write client-facing deliverables for Aurora Coffee, an independent
    coffee chain with five locations in the Pacific Northwest. The CMO
    prefers concise, plain prose with concrete metrics. Always include a
    "Risks and Assumptions" section in proposals. Never use the words
    "delight", "unleash", or "seamless".
  5. Model: pick a sensible default — usually the workspace default Claude model is right.

  6. Style: open the Style tab and upload 1–5 representative samples of past Aurora work. Click Extract. Review the resulting tone_descriptors, banned_phrases, etc.

[SCREENSHOT: agent builder with Aurora Coffee agent open on Style tab]

Using the agent in a workroom

When you create a workroom for Aurora Coffee, add the Aurora Coffee — Proposal Drafter agent to the workroom. The Composer picks the first non-system custom agent in the room as its style source. To make sure it picks the right one, add only that agent.

Naming convention

Recommend:

{Client} — {Deliverable Type}

Examples:

  • Aurora Coffee — Proposal Drafter
  • Aurora Coffee — Weekly Report Writer
  • Northwind Real Estate — Pitch Deck Outliner

Splitting by deliverable type lets you keep different formats consistent without the agent's system prompt sprawling.

Versioning

Each edit creates a new agent version. Past versions are kept for audit (bot_versions). If a new version produces worse output, revert from the agent detail page.

Per-plan caps

The number of custom agents per workspace is capped by max_bots:

PlanCustom agents
Free / Solo2
Starter3
Pro and aboveUnlimited

If you serve three or more distinct clients, plan to be on Pro. See pricing.

Common pitfalls

  • One generic "agency writer" agent for everything. Defeats the entire point. Build per-client agents.
  • Identical samples across agents. Don't reuse the same five samples for every client agent. Use samples that actually represent that client's voice.
  • Long system prompts. Keep the system prompt focused on role and policy. Voice rules belong in Style Memory, not in the prompt.
  • Forgetting to add the agent to a workroom. Without the agent in the room, the Composer falls back to the system default. Add it on workroom creation.

What's next

On this page