Intent Recognition Evolved: From Rule-Based Classification to Goal-Oriented Agent Harnesses

01 Evolution Mainline: From Rules to LLMs

Intent recognition — the core NLU task answering "what does the user want to do?" — has evolved through distinct generations, each solving the previous generation's fatal flaw:

Rules : Keyword/regex matching; brittle, no generalization.

Statistical ML : Feature engineering (TF-IDF, n-grams) with classifiers (SVM, CRF); requires heavy feature craft.

Deep Learning : Embeddings (Word2Vec, BERT) capture semantic similarity; handles paraphrase but struggles with long-tail intents.

Transformers : Self-attention models context; enables few-shot intent classification.

LLMs : Large language models perform open-ended intent understanding via prompting and in-context learning; shift from fixed-label classification to open-ended goal interpretation.

The article illustrates this progression with a diagram (not reproduced here) showing the capability ladder.

02 Two Orthogonal Expansion Axes

Beyond the mainline, intent recognition expands along two independent, combinable directions:

A. Multimodal Intent Understanding (Input-Side Expansion)

Intent signals are not text-only. The perception entrance widens to voice, image, sensor, and behavioral data.

Example : User uploads a photo of a device with a blinking red light and asks "Why is this flashing red?" The system jointly understands device model, light location, and text → intent Troubleshoot_Device_Error.

Modalities: Text | Voice | Image | Sensor/Data.

B. Intelligent Agent Action (Output/Capability-Side Expansion)

After recognizing intent, the system moves from classification to goal-oriented task execution: decompose goal, plan, call tools, complete action.

Flow : Intent → Goal → Plan → Action.

Example : "Book me a flight" → understand destination/date → compare constraints → search options → prepare booking → request confirmation.

These axes are orthogonal, not sequential: multimodal expands input, agentic expands output. They combine into "multimodal-input agent systems." The article proposes the more accurate framing: From Intent Recognition to Goal-Oriented Task Execution .

Five Dimensions of Evolution (Cross-Cutting)

A table summarizes synchronous shifts across five dimensions:

Expression Understanding : Keyword → Features → Embeddings → Context → Reasoning

Intent Space : Fixed labels → Expandable intents → Open-ended goals

Model Paradigm : Rules → ML → Deep Learning → Transformers → LLM

Input Form : Text → Text + Voice + Vision + Sensors

Output Capability : Classification → Explanation → Planning → Execution

03 Position in Agent Harness

Intent recognition is not a standalone module but a layered decision layer straddling the boundary between "input" and "decision/execution," spanning multiple gates. It converts natural-language requests into executable judgments: what task, who handles it, is it safe, is information sufficient — before planning and execution.

Location : Front of the control flow as a "gate layer," not a data layer. It outputs decision signals (classification, routing target, risk level, clarification flag), not content.

An architecture diagram (referenced in source) shows this positioning.

04 Scenario Application: Harness Conversational Interface (Feishu Group Chat)

Context

The author builds an AI Coding Harness with long execution chains, async flows, and multiple human decision points. CLI cannot support "conversation + async approval"; IM (Feishu) becomes the "human–Harness dialogue" entry.

Architecture: Channel Adapter + Four Intent Types

Channel Adapter is the key: Harness Core knows nothing of Feishu/DingTalk/Slack; it only accepts structured Command. The LLM does one thing — convert natural language to structured Command — acting as the core hub.

End-to-end flow : IM Message → Session Entry → Channel Adapter → Intent Parser (LLM) → Command → Authorization → Application Service → Harness.

Example Command :

{"command":"task.approve","taskId":"TASK-103"}

Why Not Direct Message-to-Command Mapping?

Group chats contain both "do" (actionable) and "ask" (query-only) messages. "Ask" messages have no executable object and would yield mechanical replies ("I can run task-xxx…"), yet they are high-frequency. Therefore, messages must first be routed by intent category before deciding handling.

Four Intent Categories

Q (Query) : Wants existing facts only. Side effect: none (read-only).

A (Action) : Directs action on existing object (run/cancel/approve/reject/publish/change-config/authorize). Side effect: controlled, auditable.

N (New Work) : Requires changing system behavior or deliverables; needs human code. Flow: Problem → Clarify → Spec → Task.

C (Chat/Other) : Neither query nor work (greetings, capability questions, off-topic). Side effect: none.

Implementation: Two Gates

Gate 1 — Deterministic Rules : Catch structurally regular messages (e.g., "view TASK-103", "cancel RUN-203") via regex. Saves tokens, avoids model non-determinism.

Gate 2 — Model with Context : Remaining natural-language messages go to LLM, which receives current session state. Single-message classification fails (e.g., "OK" alone is ambiguous); context resolves which task is being confirmed.

Why Rules First? Deterministic problems → code; fuzzy expressions → model.

Why Context? Model outputs structured intent conditioned on conversation state, not single-message mapping.

Gate 3 — Confirmation : High confidence ≠ auto-execute. Ambiguous or high-side-effect messages trigger clarification. Only explicit Actions (A) and confirmed New Work (N) become Commands, pass authorization, and enter Harness Core.

A diagram (referenced) illustrates the two-gate flow.

05 One-Sentence Summary

Intent recognition has evolved from "classifying predefined intents" to "interpreting open human goals in context," expanding along two independent axes — multimodal perception and agentic execution. In engineering, it becomes a layered decision layer in Agent Harness that spans the input→decision/execution boundary, converting requests into executable judgments: what task, who handles it, is it safe, is information sufficient.

Persistent Challenges (Nature of Language & Intent)

Ambiguity & Indirect Expression : Humans often express goals indirectly.

Mid-Interaction Intent Drift : Goals may change during interaction.

Implicit Context Unspoken : Critical information not explicitly given.

Personalization vs. Privacy : Adaptation must avoid over-collection.

Distinguishing Real Intent from Incidental Mentions : Correctly separate what the user wants from what they merely mention .