Classifiers - Hue

Classifiers read your production traces and answer questions about them: whether a trace needs attention, what task a request was, and how your agent behaved across many requests. They work from what your traces record, without changes to your agent, and their results are advisory: they never score a run. To score runs, use evaluators.

What Hue classifies

Classifier	What it answers	Where you see it
Findings	Whether an ended trace shows a problem: a tool error with no later success, a failed model call, no reply, or a reply that says the agent couldn’t help, missed the request, didn’t carry out a requested action, or claims an action no tool performed	The attention state and Findings filter in Traces; `search_traces` with `attention` or `finding`
Intents	Which task type each request belongs to, from the categories you define	User Intent; `get_intent_summary` and `list_intent_traces`
Request answers	How the agent behaves across requests: delegation to sub-agents, work added beyond the request, handoff scope, tool output it dropped, slow task types, what users ask for and time to answer	`get_request_answer` over the Hue MCP server
Trace checks	Project-defined, versioned checks of delegation, action scope, handoff scope and omitted tool output, with request timing	Traces and `get_trace_check_results`

Each result links to the traces behind it, so you or your coding agent can read the evidence. Findings skip traces your evaluation runs record.

When classifiers run

Classifiers are on for every project; nothing needs to be switched on in Settings.

Findings run on each ended production trace. Rule findings stay inside Hue; findings about the reply use a model.
Intents start once you define your task categories on User Intent.
Request answers start once Hue maps your project’s requests and activates judging for it. get_project_context reports whether they are available.
Trace checks need instrumentation that marks each request and its agent boundaries and an activated, tested version; see trace checks and request timing.

Several of these send bounded conversation text to a model. AI processing lists what each one sends. How you define and tune classifiers for your own questions is changing. To set up request answers or trace checks for your project, ask the Hue team.

Read results with your coding agent

Ask in plain language, for example “What needs attention in the last 24 hours, and why?” or “How often does the agent hand work to sub-agents?”. The MCP recipes give the tool sequences, and the data model defines every attention state, finding and result field.