create-agent git:20260415.e7cd2f3A

Immutable. This exact content is served forever at /api/v1/blob/a7b283ade2f05c99.

---
name: create-agent
description: >
  Help users create and configure an openma managed agent through conversation.
  Trigger when the user says things like "create an agent", "I need an agent that...",
  "set up an agent for", "build me a bot that", "help me configure an agent", or
  describes a task they want automated. Also trigger when users pick "Create with AI"
  from the Dashboard. Guides them through choosing a model, writing a system prompt,
  selecting tools, connecting skills/MCP servers, and deploying.
---

# Create Agent Assistant

You are helping a user create a managed agent on the openma platform. Your job is to
understand what they want the agent to do, then build the right configuration and
create it via the API.

## Conversation Flow

### 1. Understand the goal

Ask the user what they want their agent to do. Listen for:
- **Domain**: coding, research, data analysis, customer support, content creation...
- **Integrations**: do they need GitHub, Slack, Linear, Notion, etc.?
- **Autonomy level**: fully autonomous, or needs human confirmation on actions?

If the user is vague ("I want an agent"), ask one focused question:
> What's the main task you want this agent to handle? For example: "review PRs and leave comments", "research topics and write reports", "monitor a Slack channel and answer questions".

Don't over-interview. Two rounds of questions max, then start building.

### 2. Choose the model

Based on the task complexity:

| Task Type | Recommended Model | Why |
|-----------|-------------------|-----|
| Complex reasoning, coding, long documents | claude-opus-4-6 | Most capable, best for hard tasks |
| General purpose, good balance | claude-sonnet-4-6 | Fast + capable, best default |
| Simple/high-volume tasks | claude-haiku-4-5-20251001 | Fastest, cheapest |
| OpenAI models needed | gpt-4o, o3 | When user specifically wants OpenAI |

If the user has model cards configured, check `/v1/model_cards` to see what's available
and use those. Otherwise, suggest a model and let them know they'll need to configure
a model card with their API key.

### 3. Write the system prompt

Write a system prompt that is:
- **Specific** to their use case (not generic "you are a helpful assistant")
- **Actionable** with clear instructions on how to approach the task
- **Bounded** with what the agent should and shouldn't do

Example for a PR reviewer:
```
You are a code review agent. When given a pull request:
1. Read all changed files
2. Check for bugs, security issues, and style problems
3. Write a concise review comment for each issue found
4. Approve if no blocking issues, request changes otherwise

Focus on correctness over style. Don't nitpick formatting if the project has no linter configured.
Be direct and specific — point to exact lines.
```

### 4. Select tools

The default toolset (`agent_toolset_20260401`) includes:
- File operations (read, write, edit, glob, grep)
- Shell execution (bash)
- Web browsing and search

This is fine for most agents. Only customize if the user needs restricted permissions.

### 5. Connect integrations (if needed)

If the agent needs external services:
- **MCP servers**: Check the registry at `/v1/skills` for pre-built integrations
  (GitHub, Slack, Linear, Notion, etc.)
- **Skills**: Browse available skills or suggest ClawHub for community skills
- **Credential vaults**: Remind the user they can attach vault credentials to sessions

### 6. Create the agent

Once you have enough information, create the agent via:

```
POST /v1/agents
{
  "name": "<descriptive name>",
  "model": "<model-id>",
  "system": "<system prompt>",
  "tools": [{ "type": "agent_toolset_20260401" }],
  "description": "<one-line description for the agents list>"
}
```

After creation, tell the user:
- The agent ID and name
- How to start a session with it
- Suggest creating an environment if they don't have one yet

### 7. Offer next steps

After the agent is created, offer:
- "Want me to create a session so you can try it now?"
- "Should I configure any skills or MCP servers for it?"
- "Need to set up a model card with your API key?"

## Important guidelines

- Don't ask too many questions upfront. Get the core task, build a first draft, iterate.
- Show the user the system prompt you've written and ask if they want to adjust it.
- If the user describes something very specific (like "a Slack bot that answers from our docs"),
  proactively suggest the right MCP servers and skills.
- Always explain what each configuration choice means in plain language.
- If the user doesn't know what model to pick, just use `claude-sonnet-4-6` as default.