Most investigative work is a long chain of small steps: run a search, read the result, decide what it means, choose the next search. Each step is simple. Keeping all of them straight over hours, without losing a lead or repeating one, is the hard part.
Durandir Agent does that part. You describe the objective in plain language, and it turns it into a plan: what to look for, in which order, with which tools. Then it carries the plan out, step by step.
It works with the same tools you would use by hand, and everything it does is recorded in the case.
It asks before it acts
The agent does not run unattended by default. When a request is ambiguous, it asks what you mean instead of guessing. When a step touches something sensitive, such as contacting a system, spending a paid lookup or acting under your identity, it stops and waits for your approval.
Permissions are set per case. You decide which tools the agent may use, which targets are in scope and how far it may go before checking in. It cannot widen those limits on its own.
You can stop it, redirect it or take over at any point. The work done so far stays in the case.
Checked against the evidence
A language model can be confident and wrong. Durandir Agent is built on the assumption that it will sometimes be, so it does not take its own word for anything.
Every conclusion has to cite the evidence that supports it: the exact captured text, the step that produced it. A separate review compares the claim with the citation. A claim that the evidence does not support is marked as unverified and cannot be used to justify the next step.
What reaches you is an answer with its sources attached, and a clear line between what was confirmed and what is still an open lead.
Your choice of model
The agent is not tied to one AI provider. It can run on the provider your organisation already uses, and it keeps its task memory in the case, not with the provider.
If a provider is unavailable, the task pauses with its state saved and resumes where it stopped.