Build Custom Agents

You can create a custom agent in under 5 minutes. Describe what you want in plain English, and Arfiti generates the complete configuration — system prompt, tool permissions, and schedule. No coding required.

In this guide, we will build an AR Aging Reporter that monitors overdue invoices every Monday and emails a summary to the finance team.

---

Step 1: Open the Agent Creator

Go to Agents > Agents in the sidebar, then click Create Agent in the top right corner.

You will see two tabs: AI Generation (Recommended) and Manual Entry. We will use AI Generation.

---

Step 2: Describe What You Want

In the Describe Your Agent textarea, write a clear description of what the agent should do. Be specific about:

• What data to look at (AR invoices, bank transactions, etc.)
• What action to take (send email, create report, post transaction)
• When to run (every Monday, when an email arrives, on demand)
• Who to notify (email address, team)

Here is the description we used for our example:

Create an agent that monitors overdue AR invoices every Monday morning. It should find all invoices overdue by more than 30 days, group them by customer with total amounts, and send a summary email to hello@ar-ti-fi.com. Include the number of overdue invoices and total outstanding amount in the email subject line.

Tip: The more specific your description, the better the generated configuration. "Monitor invoices" is vague. "Find AR invoices overdue by more than 30 days, group by customer, email summary to finance team" gives the AI everything it needs.

---

Step 3: Generate with AI

Click Generate with AI. The system runs a two-pass analysis:

1. Pass 1 (fast): Analyzes your description to identify which documentation, tools, and workflows are relevant
2. Pass 2 (thorough): Generates the complete agent definition — system prompt, tool list, configuration

This takes about 30-45 seconds and costs approximately $0.08.

---

Step 4: Review the AI Analysis

When generation completes, you will see an AI Analysis section showing what the AI decided:

• Suggested Type: The identifier for your agent (e.g., ar_aging_monitor)
• Complexity: How complex the AI considers the task
• Context Used: Which documentation and system data informed the generation
• Cost: The token cost for this generation

Claude.ai — Screenshot placeholder: Asking Claude to help design an agent — e.g., "I want an agent that sends me a weekly summary of overdue invoices" and Claude suggesting the description to use

---

Step 5: Review the Generated Configuration

Scroll down to see the complete configuration the AI generated. This is the most important part — review it before creating.

The key fields:

• Agent Type: A unique identifier like ar_aging_monitor (auto-generated)
• Name: Human-readable name shown in the dashboard
• System Prompt: The instructions Claude will follow when the agent runs. This is the heart of the agent — it tells Claude exactly what to do, step by step
• Max Concurrent Runs: How many copies can run at once (1 for most agents)
• Timeout: How long before the agent is stopped if it has not finished

Scroll further to see the tools and configuration:

• Allowed Tools: Which MCP tools the agent can use. The AI selected list_entities, get_entity, send_email, store_memory, and recall_memories for our reporter
• Configuration: Custom parameters like the overdue threshold (30 days), recipient email, and schedule

You can edit any field before creating. If the system prompt needs adjustment, change it directly in the textarea.

---

Step 6: Create the Agent

Click Create Agent Definition. You will be redirected to the definitions list where your new agent appears alongside the default agents.

Your agent is now registered but not yet running. You need to trigger it.

---

Step 7: Trigger Your Agent

Option A: Schedule and Run Now (Recommended)

Go to Agents > Scheduled Jobs and click Create Job. Select your agent, set a cron expression (e.g., 0 9 * * 1 for every Monday at 9am), choose a timezone, and save.

To run the agent immediately, click the play button (▶) at the end of the row. The agent will start within 30 seconds.

Option B: Ask Claude

Claude.ai — Screenshot placeholder: Telling Claude "Run the ar_aging_monitor agent for entity 12" and Claude confirming the agent has been triggered

---

Step 8: Monitor the Run

Go to Agents > Sessions to watch your agent execute. It should appear within 30 seconds of being triggered.

The sessions list shows:

• Status: Running (yellow), Completed (green), or Failed (red)
• Agent: Which agent ran
• Duration: How long it took
• Cost: API cost for LLM agents ($0 for code agents)

---

Step 9: Check the Results

Click any session to see the full execution trace — every tool call, every decision, every output.

The Agent Summary shows what the agent accomplished. The Progress Log tab shows the step-by-step execution. If something went wrong, this is where you find out why.

Claude.ai — Screenshot placeholder: Asking Claude "Did the ar_aging_monitor run successfully?" and Claude summarizing the results

---

If Something Goes Wrong

---

Tips for Writing Good Descriptions

1. Be specific about the data source: "AR invoices overdue by 30+ days" not just "overdue invoices"
2. Name the action: "send a summary email" not just "report"
3. Include the recipient: "email to hello@ar-ti-fi.com" or "notify the finance team"
4. Set the schedule: "every Monday morning" or "when a new bill is posted"
5. Mention thresholds: "more than 30 days overdue" or "amounts above 1,000 EUR"

---

Choosing the Right Tools

When creating an agent manually or editing a generated one, these are the most commonly used tools:

Fewer tools = lower cost. An agent with 7 tools costs ~$0.05 per run. An agent with all 664 tools would cost ~$0.30 per run.

---

Next Steps

• Manage Agents — monitor agent activity, view memory, check the event queue
• Agent Orchestration — coordinate multiple agents with sagas and downstream rules