Coding Agents in Practice (4/5) — Multi-Agent Patterns: Orchestrator and Specialist Separation

AI ์ฝ”๋”ฉ ์—์ด์ „ํŠธ ์‹ค์ „ (4/5) — ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ํŒจํ„ด: ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ดํ„ฐ์™€ ์ „๋ฌธ๊ฐ€ ๋ถ„๋ฆฌ

When a single LLM session does everything, the context breaks. Role separation fixes it.

ํ•ต์‹ฌ ์š”์•ฝ

  • Multi-agent is not "let's use more model" — it's "let's separate context"
  • Primary sources: Anthropic Claude Code subagents docs; OpenClaw / Hermes operational experience
  • Three core patterns: orchestrator + specialists / pipeline / collaboration
  • Cost is not a simple multiplication — context savings and reduced retries often make multi-agent cheaper overall
  • Common pitfalls: permission leakage, context contamination, debugging difficulty

1. Why one agent doing everything fails

When a single LLM session handles design + implementation + testing + review + docs, three failure modes appear:

  • Context contamination: details from one task pollute another. Debug-time scratch code becomes a PR-review baseline.
  • Context limit overflow: even 200K tokens fill quickly under heavy load. Auto-compaction shaves the reasoning behind decisions.
  • Permission asymmetry: the same session has publish + credential access + code change authority. One mistake has wide blast radius.

Solution: separate agents by task type and have one orchestrator on top.

2. Orchestrator + specialists pattern

The most common topology.

        [Orchestrator = main]
            │
   ┌────────┼────────┬────────┐
   ▼        ▼        ▼        ▼
 [Edit]  [Test]  [Review]  [Document]
 executor  tester  reviewer   writer

Orchestrator's role: - Receive a request and decide which specialist to call. - Absorb specialist results in summary form (not full output). - Decide the next step.

Specialist's role: - Work deeply in their specialty. - Return a condensed report — protecting the orchestrator's context.

Core principle: orchestrator sees the whole picture, specialists see depth. Two kinds of context, neither overloaded.

3. Three topologies

3.1 Supervisor (hub-and-spoke)

  • One orchestrator + N specialists, as above.
  • Simple, predictable, easy to debug.
  • Limit: when specialists need to talk directly, every hop goes through the orchestrator.

3.2 Pipeline (serial)

  • Specialist A → B → C, output of one becomes input of the next.
  • Strong fit for clear-stage work — data transform, validation, publish.
  • Limit: a failure at one stage halts everything.

3.3 Collaboration (mesh / consensus)

  • Multiple specialists see the same input in parallel and produce answers; results are merged or voted.
  • Good for code review where different lenses (security / performance / readability) matter.
  • Most expensive: same tokens processed N times.

Selection rule: clear stages → pipeline; clear division of labor → supervisor; different lenses → collaboration.

4. Claude Code's subagents — Real implementation

Claude Code calls subagents through the Agent tool, with subagent_type selecting the specialist type and a self-contained prompt.

Prompt-writing principles (where most teams stumble): - The subagent does not know the main agent's context. Put what / why / what's been tried all into the prompt. - Specify the output shape: "report in 200 words," "JSON result," "return only the URL," etc. - Don't delegate understanding: instead of "analyze and decide," say "in file X line Y, change Z to W." The main agent must understand first, then delegate.

Parallel vs serial: - Independent investigations can fan out in parallel. - Dependent work runs serial. - Parallel is faster but the total token cost is roughly the same.

5. Cost — Not a simple multiplication

"Multi-agent = N× cost" is half-true.

What costs more: - Each subagent re-loads its own context (CLAUDE.md, rules, tool defs). - Orchestrator-specialist message round-trips themselves cost tokens.

What costs less: - The main context stays light, so less compaction — and compaction triggers re-asks. - Specialists can run on smaller models (e.g., Haiku) when the work allows. - Fewer retries — when the main is contaminated, decision quality drops and whole tasks restart.

Heuristic: simple 1–2 step tasks → single agent, almost always cheaper. 5+ step or mixed-specialty work → multi-agent often wins on total cost.

6. Four common pitfalls

6.1 Cross-contamination

  • The main agent ingests a specialist's full output and stores it in memory → contamination.
  • Fix: specialists return summaries only; the main agent trusts only the summary.

6.2 Permission leakage

  • Every agent gets every tool → wide blast radius from one mistake.
  • Fix: per-agent tool whitelists. Only a designated agent can publish.

6.3 Debugging difficulty

  • "Which agent went wrong, and where?" is hard to trace.
  • Fix: every agent call logs a correlation ID + result summary. At session end, archive logs separately.

6.4 Over-engineering

  • Inserting multi-agent into trivial work.
  • Fix: a hard rule like "3+ step tasks enter multi-agent flow." Below that, single agent + self-check.

7. At a glance

Pattern Best for Strength Weakness
Orchestrator + specialists (supervisor) Clear division of labor Simple, easy to debug Specialists can't talk directly
Pipeline Clear stages Trace easy One stage fails → whole halts
Collaboration (mesh / consensus) Different lenses Diversity Most expensive
Single agent ≤2 step tasks Cheapest Context contamination risk

Design rule: task types diverge → agents diverge. Otherwise, single agent.

Next up

Part 5/5: Coding Agent Cost Management — Tokens, Caching, Routing. If multi-agent splits structure, the next post splits cost.

References

  • Anthropic, Subagents in Claude Code — code.claude.com/docs/agents (verified 2026-05-05).
  • Anthropic, How We Built Our Multi-Agent Research System — anthropic.com/research/built-multi-agent-research-system (verified 2026-05-05).
  • The "What Is Harness Engineering?" series — theoretical background.

This is part 4/5 of the Coding Agents in Practice series.

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

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

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ 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"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š 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"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š 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. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

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

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š 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. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

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

← 6/9 Neural Networks ๐Ÿ“š 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. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

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

← 1/13 Current Structure Inve… ๐Ÿ“š 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 ๐Ÿ“š 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 »