Which agents use live web search and which rely on model knowledge
Not every specialist agent goes online. Two of the six roles use live web search; the rest answer from the model's own knowledge. This matters when you need current facts versus general reasoning.
Roles that use live web search
The research and reasoning sub-agents use live web access when they run. This is implemented through OpenRouter's bounded web plugin, capped at a small number of results (max 3). If a web call fails, the system falls back to a no-web call so the agent still returns an answer rather than erroring out.
Roles that rely on model knowledge
The other four roles — design, document, patent, and manufacturing — rely on the model's built-in knowledge rather than searching the web. Their outputs reflect what the model already knows, shaped by their role-specific system prompts.
What this means in practice
- Need current market data, competitors, or recent facts? Route the work through research or reasoning so it can consult the web.
- Need a design spec, a polished document, patent-style claims, or a sourcing plan? Those roles draw on model knowledge, so treat time-sensitive specifics with appropriate caution.
Flows behave the same way
Flow steps run the same role logic as Camelia's ad-hoc sub-agents, including live web for research and reasoning steps. So if you build a multi-step flow that starts with a research step, that step gets the same bounded web access it would in a direct delegation.
Keeping web access scoped to research and reasoning keeps results focused and avoids unbounded crawling, while still letting the investigative roles ground their answers in current information.