Durandir Agent is our working answer to that question: it plans, coordinates the tools of each product and checks its results against the evidence before it reports them. Research in this area decides how each of those steps works.
What we study
- Planning: breaking an investigative or security objective into steps that can be checked, and revising the plan when a step fails.
- Tool use: letting an agent run real tools, from searches to scanners, inside a scope and with a record of every action.
- Verification: checking each claim against the evidence that supports it, and saying so when the evidence is not there.
- Guardrails: safeguards that act before anything runs, judge the request rather than its wording, and cannot be talked around.
Why it matters
An agent that cannot show its work is of little use in investigation or security. We want agents whose every conclusion can be traced back to a source and repeated by someone else.