Every agent needs a different bar for risk and quality depending on what it actually does. These are the archetypes Prefactor's case studies and content are organised around.
Agents that resolve or triage customer conversations directly — chat, ticket, or in-app support — and need a clear, auditable point where they hand off to a human instead of guessing.
Conversational voice agents plus the background agents and after-call workflows they trigger — each layer carrying a different risk profile that needs to be tracked separately but linked to the same call.
Agents built for internal operations rather than customers — typically shipped fast with a lower risk bar, but still needing a kill switch and an audit trail once they touch real systems.
Agents that write, review, or ship code changes — from autocomplete-style assistants to autonomous PR-opening agents — where the controls that matter are scoped repo/tool access and a reviewable trail of what changed and why.
Agents that gather, synthesize, and report on information across documents, APIs, or the web — where the main governance question is whether an answer is actually grounded in what it retrieved.
Agents specialized in retrieving from knowledge bases, vector stores, or internal APIs to ground another agent's response — typically read-only, but still needing data-access controls and retrieval-quality checks.
A supervising agent that decomposes work across specialist sub-agents — where a single misrouted step can cascade, so tracing the full chain back to one instance matters more than any individual agent's output.
Agents that run without a human waiting on the other end of the conversation — scheduled jobs, event-triggered workflows, long-running tasks — where a kill switch and drift detection matter more than real-time latency.
Prefactor helps teams observe, evaluate, and improve their AI agents in production — across every framework and provider.