WardrikResearch (wardrik_research v0.1.0-rc.1)

Deep research for agents, as the deep_research tool.

It splits a topic into angles, researches each one in its own subagent at the same time, and combines the findings into one report.

WardrikCore.start_agent(
  agent_id: "acme/analyst",
  model: "anthropic:claude-haiku-4-5",
  capabilities: [{WardrikBrave, brave: MyApp.Brave}, WardrikResearch]
)

Each researcher runs on the agent's model, middleware and tools, except deep_research. Give the agent a search tool, or research only knows what the model knows. What researchers spend counts toward the agent's budgets, and what they read labels the answer.

Options

{WardrikResearch, angles: [%{angle: "pricing", prompt: "Think like a pricing analyst."}]}

• :angles - the angles a topic is split into, each %{angle: name} with an optional :prompt added to that researcher's instructions. Five general angles unless set.
• :max_concurrency - how many researchers run at once. 5 unless set.
• :timeout - how long one researcher may take, in milliseconds. 8 minutes unless set.
• :max_rounds - how many model calls one researcher may make. 12 unless set.

The agent may pass its own angles for a single call. A failed or timed-out angle is reported as a gap, and the rest are still combined. The call is an error only when every angle fails.