Multi-Agent Routing Design — Classifier Strategy, Provider Abstraction, Cost Lifecycle

The most expensive mistake in routing layer design is deferring model selection until runtime.

Key Takeaways

  • The request classifier strategy is the primary determinant of routing layer quality.
  • Multi-provider abstraction is not about unifying interfaces — the core is fallback chain design.
  • Without integrating cost lifecycle into the model selection loop, a routing layer adds complexity without cost optimization.

Design Background

These are structural lessons from designing a multi-provider LLM routing layer (Router_Control) in Node.js + TypeScript. Target providers: OpenAI, Claude, Ollama, oMLX. The goal was automatic distribution of requests to the appropriate model based on request type.

This post focuses not on a success story, but on which design decisions determined outcomes.

Body

1. Classifier Strategy — The Real Quality Determinant of Routing

The first design decision in a routing layer: how to classify requests.

The initial design was keyword-based classification. If the request text contained keywords like "code," "translate," or "summarize," it routed to the corresponding model. Fast and simple to implement.

The problem: keywords do not represent intent. "What do you think of this code?" is an interpretation request, not code generation. Keyword-based classification cannot distinguish between the two.

The approach that worked:

interface ClassifierResult {
  taskType: 'code-gen' | 'code-review' | 'translation' | 'embedding' | 'general';
  complexity: 'low' | 'medium' | 'high';
  latencyRequirement: 'realtime' | 'batch';
  costBudget: 'local-only' | 'cloud-ok' | 'premium-ok';
}

Represent the classification result as a multi-dimensional struct rather than a single label. Combining task type, complexity, latency requirements, and cost budget is what makes routing decisions meaningfully precise.

The classifier itself can be delegated to an LLM, but that incurs classifier call costs. A two-tier structure — rule-based for simple requests, LLM classifier only for complex ones — was the cost-accuracy balance point.

2. Multi-Provider Abstraction — The Interface Unification Trap

The standard approach for multi-provider routing is to define a common interface:

interface LLMProvider {
  complete(prompt: string, options: CompletionOptions): Promise<CompletionResult>;
}

This abstraction is clean at the code level. In production, it has a structural problem.

Each provider fails differently. OpenAI fails with rate limit 429. Ollama fails with a timeout when the model isn't loaded. oMLX terminates the process on memory exhaustion. Hiding these differences behind a common interface means fallback logic cannot distinguish failure causes.

A structurally sounder approach:

interface ProviderError {
  type: 'rate-limit' | 'model-unavailable' | 'resource-exhausted' | 'network';
  retryable: boolean;
  fallbackSuggestion: ProviderId | null;
}

Type failures per provider, and branch the fallback chain by failure type. Rate limits retry to another endpoint on the same provider; unloaded models immediately pass to the next provider; resource exhaustion downgrades to a lighter model.

The objective is not interface unification — fallback chain design by failure type is what matters.

3. Cost Lifecycle — Meaningless Without Integration Into the Selection Loop

If a routing layer aims for cost optimization, cost information must exist inside the model selection loop.

The problem with the initial design: cost tracking was logging. Token consumption was recorded and displayed on a dashboard, but that information did not influence model selection for subsequent requests. Monitoring and control were decoupled.

Integrating cost lifecycle into the routing loop:

interface CostBudgetTracker {
  sessionBudget: number;           // total session budget
  consumed: number;                // consumed so far
  remainingRatio: number;          // remaining ratio
  nextModelConstraint(): CostBudget; // model constraint based on remaining budget
}

Remaining budget above 50%: premium models permitted. Below 20%: local-only enforced. This constraint combines with the classifier result to determine the final model.

Without this structure, the routing layer does not optimize cost — it only adds routing complexity.

4. OpenClaw–Hermes Integration History — Factual Record

The routing layer was designed for integration with the OpenClaw (OC) multi-agent platform. The structural transitions that occurred in this context are recorded as-is.

Sequence: OC stable operation → Hermes (HM) first migration attempt → token explosion → OC regression → OC redesign → HM retry (currently in validation).

Router_Control entered Suspended status due to priority re-adjustment during the OC→HM transition. This was not a technical failure. The token explosion issue was on the HM side; the routing layer itself operated at prototype level.

Once HM retry validation completes, resuming the routing layer based on the design patterns covered here — multi-dimensional classifier, typed fallback chains, cost loop integration — is under consideration.

Related PR: harness git PR#12497.

5. The Design Value of Suspended Status

Marking a component as "Suspended" in an ontology serves a purpose beyond simple status indication.

  • Context preservation: The design rationale for why it started, what it achieved, and why it stopped is preserved.
  • Reduced resumption cost: Resumption does not require redesigning from scratch — it can restart from the previous decision point.
  • Decision reference: When the same problem is encountered again, prior decision context is immediately retrievable.
  • Ontology honesty: An ontology with only Active and Completed hides suspended components. Explicit Suspended status exposes the system's actual state.

In multi-agent systems, each component carries a maintenance cost. Continuously evaluating value against cost — and honestly transitioning low-value components to Suspended — is how overall system efficiency is maintained.

Lessons Learned

Designing the classifier with a single label: Initially routing on taskType alone. Routing accuracy improved meaningfully when complexity and cost budget were added as separate dimensions.

Fallback logic that did not distinguish failure causes: Any failure was simply passed to the next provider. Only after typing failure modes did the responses to rate limits and resource exhaustion diverge.

Cost tracking that stayed at logging: Costs were recorded but never fed back into the model selection loop. Separating monitoring from control produces visibility, not optimization.

Conclusion

The real quality of a multi-provider routing layer depends not on the cleanliness of interface abstraction, but on three design decisions: multi-dimensional classifier, typed fallback chain by failure mode, and cost loop integration. Without these three inside the selection loop, a routing layer is middleware that only adds complexity.

Router_Control is currently Suspended, but the design patterns remain valid. Resumption on this foundation is planned after HM validation completes.

댓글

댓글 쓰기

이 블로그의 인기 게시물

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… 📚 Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere 핵심 요약 This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

이미지
← 8/9 Deep Learning Architec… 📚 Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

이미지
Series overview: Series index 📚 Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. 핵심 요약 An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

이미지
← 7/9 Deep Learning Training 📚 Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. 핵심 요약 CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks 📚 Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. 핵심 요약 SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… 📚 Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent 📚 Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »