Academy · Platform · Agents
Agent varieties
In one line. Same anatomy, different jobs — the four working styles an agent takes on this platform, and how to pick the right one for the task in front of you. You’ll be able to. Name the variety a problem calls for before you open the editor, so the persona, knowledge and wiring follow naturally.
Why varieties matter
Every agent shares the same anatomy — model, persona, knowledge, capabilities, guard rails. What changes is the job contract: what comes in, what must come out, and where the output lands. Get the variety right and the rest of the configuration writes itself. Get it wrong and you’ll fight the persona for a week.
The four working styles
1. Answering — the expert you ask
In: a question. Out: a grounded, cited answer. Lands in: chat.
The Agentic Search worker. Attach Knowledge collections, demand citations, instruct refusal when the documents don’t contain the answer. Success is measured in trust: a user who clicks a citation and finds exactly the supporting passage comes back tomorrow.
Build it in S3 · Your first 30 minutes; harden it in the Search pattern.
2. Extracting — the analyst who fills forms
In: a document. Out: schema fields with per-field confidence. Lands in: the collection’s records, visible in the workspace.
The workhorse of every Workflow. The schema is its real contract — a well-described field (“total invoice amount including tax, as printed”) outperforms any persona flourish. Confidence per field is what makes the human-in-the-loop gate possible: high goes through, low goes to review, and reviewer corrections become feedback.
An extraction agent is wired to its collection as the intake/processor — see A3 for the attachment step.
3. Deciding — the officer who rules on cases
In: an assembled record (often already extracted). Out: a decision with reasons — approve, hold, escalate, route. Lands in: the lifecycle stage, the record, the audit trail.
The variety that earns governance. Persona states the policy; guard rails state the boundaries; the lifecycle carries the verdict; the trace preserves the why. Decisions that matter route through four-eyes approval — the agent recommends, a person disposes, and over time the straight-through share grows as trust accumulates.
4. Generating — the writer who produces artifacts
In: source records plus an instruction. Out: a produced artifact — a summary, a findings report, a client letter, a decision package. Lands in: the document’s summary pane, an export, or a downstream system via the Runtime API.
Generation quality is grounding quality: a summarizer attached to the settled fields writes accurate summaries; one attached to nothing writes fiction. Keep the persona’s output contract exact — length, sections, tone.
Choosing, quickly
| The sentence in your head | Variety | Start |
|---|---|---|
| “People keep asking us…” | Answering | sol-search |
| “Someone reads each one and types the values into…” | Extracting | B3 · Invoice settlement |
| “Then a supervisor approves or kicks it back…” | Deciding | G2 · Human-in-the-loop |
| “Then we write it up for…” | Generating | A3 · Your first agent |
Real processes chain them: extract, then decide, then generate. That chain is exactly what teams and meshes exist for — classify → validate → enrich → approve → summarize is four varieties holding a baton.
Where to go next
- The shared machinery underneath: A1 · Agent anatomy.
- Chaining varieties into one engine: A11 · Teams & mesh.
- Proving a variety performs before it ships: A12 · Crew & evaluations.
Prefer learning inside the product? The same academy lives in the platform's Learn menu — every screen links to the chapter that explains it.
See the platform live