An investigator's time disappears into coordination: deciding what to check next, running the same search five ways, keeping track of which thread is still open, and remembering what the third document said. The Engramite Agent takes that layer of work.
It interprets the request, selects the capabilities the task needs, and coordinates the work using the context already in the case. It is the investigation core of the workspace — and it is designed to operate inside the analyst's judgement, not around it.
What the agent does
Reads the request the way a colleague would. You describe the question — a subject, an allegation, an asset, a scope — in plain language. The agent turns that into a plan with steps.
Chooses the right capability for each step. Open-source collection, document and image analysis, relationship analysis, reporting: it picks what fits, rather than making you assemble the sequence by hand.
Keeps everything attached to the case. Each step's result lands in the case file, connected to the source it came from and the request that prompted it.
Narrates the work as it happens. You can follow the plan step by step while it runs, rather than receiving a verdict from a black box.
Three levels of latitude
The setting that matters most is how much the agent may do without asking:
- Manual — every tool run waits for your confirmation.
- Supervised — routine steps proceed, sensitive ones stop and ask.
- Automatic — the plan runs end to end and you review the result.
The same engine does the work in all three modes. What changes is where the confirmations sit — which is exactly the control an investigator needs when the subject is a named person.
What it deliberately does not do
It does not declare conclusions. A promising match is presented as a lead. Turning a lead into a finding is a review decision, and the workspace keeps those two things apart. See accepted and rejected evidence.
It does not infer identity from a name. Shared surnames, cities and professions are treated as possible confusion between people until something exact links them. Read the homonym problem.
It does not hide its sources. Every claim it surfaces points back at where it came from.
Why "agentic" is not the interesting part
Plenty of tools now run an LLM in a loop. The question that decides whether such a system is usable in real casework is narrower: can you tell what it did, can you stop it, and can you defend the output to someone who was not in the room?
That is the standard the agent is built against — planning that is visible, execution you control, and findings that carry their provenance.
Frequently asked questions
Does the agent replace the analyst? No. It removes coordination work. The judgement calls — what counts as evidence, what the finding means — stay with the investigator.
Can I see why it did something? Yes. The plan, the steps and the results are visible while it works and afterwards in the case file.
What if it gets a step wrong? You reject the result and it does not become part of the case. Rejected material stays visible as rejected, rather than silently disappearing.
Read next: Engramite Graph or what Engramite is.