Agents runs tool-calling AI agents on your model endpoint and records each step of each run.
A citizen or staff request needs several steps and several lookups.
A chat answer does not show the steps behind it.
Teams need to test an agent before it meets a citizen.
A user or a system sends a task to a published agent.
Calls your model. Runs each tool the model asks for. Stops at the step limit.
The answer, plus a record of each step and each tool call.
Agents run on the model endpoint you provide. Anar supplies no models.
Model calls and tool calls, with a step limit and a tool timeout. The default limit is 10 steps.
Get the result at the end, or stream each step as it happens.
Run history, each step, each tool call and metrics for each agent.
Traces do not record the tenant, the user or the cost yet.
agent templates with 26 tools.
Count from the product source, November 2026.
- Citizen services: citizen inquiry and permit processing.
- Staff work: HR policy, document review and procurement.
- Language work: translation and meeting summary.
- The template tools return sample data. Connect your own systems before live use.
The agent list lives in memory. A restart loses the agents you created.
An agent you build has no tools yet. Only templates have tools.
A permission list exists. The loop does not check it yet.
Any signed-in user can run any published agent.
| Capability | Status | Note |
|---|---|---|
| Agent loop, streaming, step traces | Shipped | Step limit and tool timeout. |
| Seven templates | Partial | Tools return sample data. |
| Agent builder and storage | Partial | No tools on custom agents. A restart loses them. |
| Tool permissions, run access | Partial | Not checked at run time yet. |
| Tools for other Anar products | Partial | Coded. Not attached to any agent. |
| Guard filter, approvals, connectors | Roadmap | Planned. Not built. |
- Agent loop with step limits
- Run and stream modes
- Step traces and metrics
- Arabic in every template
- Arabic operator console
- Anar Guard check before and after each run
- Connectors to your systems
- Answers grounded in tool results
- Tool grants you can review
- Tool calls on behalf of the user
- Agent versions
- Human approval steps
- A test gate before publish
Agents can send each model call through the gateway when you set its address.
Agents can send their events to the Guard audit trail. Guard does not filter runs yet.
A one-month pilot: we deploy in your environment and run it with one team.
anarlabs.ai/contact