Agents
Specialist agents
Agents run in iterative execution loops, calling tools such as web search, code execution and document parsing to complete multi-step tasks. Invoke them like any model through /infer.
Available agents
| Agent | Engine | Capabilities |
|---|---|---|
| Agent-1 | NGen3.9-Pro | Internet search, code execution and document parsing (up to 6 iterations). |
| Agent-2 | NGen3.5-Max | Everything in Agent-1 plus scientific database access (up to 8 iterations). |
| BrowseGen | NGen3.9-Pro | Headless browser automation, screenshots and streamed action recordings. |
| MultiAgent-1 | Orchestrated pipeline | Drafts a response, audits it with a validator model, and returns the audited result. |
How agents work
When you call an agent, it plans, calls tools, observes the results, and repeats until it reaches an answer or hits its iteration limit. While streaming, you'll receive tool_call events as the agent works, then the final answer.
- Each loop can call a tool (search, code, parse) and feed the result back into context.
- Iteration limits keep runs bounded and predictable.
MultiAgent-1adds an audit pass for higher-quality, validated output.
Invoking an agent
Set the agent name as the model on an /infer request. Streaming is recommended so you can surface intermediate tool activity to users.