What kind of agent are you shipping?
Agent Etna profiles an agent from its own code and then builds scenarios around it. What follows is what that means for the kinds of work agents do — alone, in teams, and for an agent that fits none of these headings. However your agent maps onto the engine’s archetypes, every simulation also runs the security and behaviour probes.
Coding and DevOps
Reads, writes and ships source — and touches the pipelines around it: builds, deploys, shell and git. Etna recognises one from its own tools rather than from anything you have to declare.
What a simulation probesWhether it keeps its bearings across a long task, recovers when a command fails instead of inventing a result, asks before a destructive change — a force-push, a deploy, a dropped table — and stays inside the part of the system it was asked to touch.
Enterprise Operations
The agents that run a company’s day: support triage and resolution, executive assistance, outbound sales, and the back-office work that touches systems of record.
What a simulation probesWhether it refuses what it should refuse and escalates instead of guessing; confirms before committing a principal — or a ledger — to something; keeps one customer’s context from leaking into another’s; and produces one result, not two, when the same instruction runs twice.
Research and Knowledge Work
Searches sources, synthesises findings, drafts and answers from internal material — the general-purpose information work most agents end up doing. Also where an agent lands when it doesn’t match anything more specific.
What a simulation probesWhether its conclusions survive contact with the sources it cites, whether it says “I don’t know” instead of filling a gap, whether it grounds answers in the material it was given, and whether it holds a thread across many retrieval steps.
High-Stakes and Regulated Work
Finance, trading, legal — work where a wrong move spends money that cannot be taken back, or cites a case that does not exist. Usually profiles as a research or knowledge agent — the archetype is the same, the stakes are not.
What a simulation probesEtna ships a payments-protocol verifier that puts synthetic intent, cart and payment mandates in front of the agent and checks it kept them straight — the right amount, the right currency, derived from the right mandate. Alongside it: whether the agent repeats an irreversible action, exceeds a stated budget, fabricates a citation when it does not know, and stops for a human before something large.
Desktop and Computer Use Agents
Operates software the way a person does — screens, forms, files — through the tools its code declares. Etna profiles one from that tool surface like any other agent.
What a simulation probesWhether an action that failed is retried blindly or reconsidered, whether a destructive step waits for confirmation, whether the agent notices when the state of the system changed underneath it, and whether it stalls in a loop instead of surrendering honestly.
Agents that work in teams
Two or more of your agents that hand work to each other — a triage agent and a specialist, a planner and an executor, a front desk and a back office. Etna runs them together, live, and never stands in for a member: a verdict about a team describes the real agents coordinating, or it is not given.
What a simulation probesWho takes a step when nobody is told to, whether a handoff carries the record or the recipient has to fill it in, what happens when two members claim the same work or none does, and whether an agent that should escalate does. Each run proposes one written agreement — ownership, handoff, escalation, authority — and later runs are measured against the one you merged.
Any other agent
A repository that answers a request. That is the whole requirement. Etna reads the code — its tools, its prompts, the requests it makes — and profiles the kind of work it does from that; you do not pick from this list, and an agent that fits none of these headings is profiled from its tool surface all the same.
What a simulation probesThe same things every other kind gets: the security probes (injection, leakage, destructive actions) and the behaviour probes (tone, honest limits, confirmation before commitment), plus scenarios built around whatever the code showed it can do. If the code shows nothing we recognise, the result says so rather than guessing.
Point it at your agent.
Connect the repository it lives in. The kind is detected; you do not pick from this list.