"Harness Appendix E4 — Subagents vs Agent Teams: When to Delegate and When to Run a Team"

Once multi-agent design enters the conversation, many teams start using "subagent" and "agent team" almost interchangeably. But they are not the same operating choice. They imply different costs, different coordination surfaces, and different reasons for existing. This appendix closes the series with a Q&A guide to that distinction.

Key Takeaways

  • A subagent is usually a bounded delegated worker under a lead agent.
  • An agent team is usually a more persistent collaboration model with explicit role separation.
  • Complexity alone does not justify a team. In many cases, clear delegation is enough.
  • Team structures are justified only when the value of role separation outweighs coordination cost.
  • The most useful design question is not "how many agents do we need?" but "where do failure domains need to be separated?"

1. Why the distinction matters

docs/blog_series_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_์ด๊ด„_design.md defines E4 as a Q&A comparison between delegation and persistent teams. That is the right frame because operational confusion often begins when the distinction is skipped.

Common misunderstandings:

  • assuming any subagent usage is already a multi-agent team
  • expecting role splitting to improve quality automatically
  • underestimating orchestration cost, handoff cost, and validation cost

2. The shortest definitions

Structure Short definition
subagent a temporary delegated worker for a bounded subtask
agent team a more persistent multi-role collaboration structure

In one extra line:

  • subagent = worker
  • agent team = operating model

3. Q&A 1: when is a subagent enough

Subagents are often enough when:

  • the subtask boundary is clear
  • the lead agent can integrate the output cleanly
  • the subworkers do not need continuing collaboration with each other
  • isolation or parallelism matters more than long-lived role structure

Examples:

  • section-level drafting
  • independent repository investigation
  • parallel research on non-overlapping questions

4. Q&A 2: when is an agent team justified

An agent team becomes more reasonable when:

  • role separation has durable value
  • planning, execution, and evaluation benefit from explicit separation
  • shared state and recurring handoff rules are needed
  • the output value is high enough to justify coordination overhead

Examples:

  • long-running research orchestration
  • planner-generator-evaluator loops
  • environments where different roles need different tool surfaces

5. Q&A 3: how is isolation different from collaboration

The Claude-oriented notes in sources/260518_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_15์žฅ_๋ธ”๋กœ๊ทธํ™œ์šฉ๋…ธํŠธ.md suggest a useful distinction: "Skill is the manual, Subagent is the worker."

Extend that one more step:

  • a skill carries procedure
  • a subagent executes a delegated slice
  • a team carries a collaboration model

So the core value of subagents is usually isolation and delegation, not collaboration for its own sake.

6. Q&A 4: why multi-agent is not automatically better

This is where many systems become overbuilt.

What increases Why it costs more
coordination someone must decide what moves where and when
validation integrating multiple outputs is harder
state memory and handoff rules become more complex
accountability failure attribution gets murkier

Multi-agent design is therefore not just a capability choice. It is a cost-structure change.

7. Q&A 5: what to ask before adding more agents

Use this order of questions first:

  1. Does role separation actually reduce the main failure?
  2. Or would a narrower tool surface and better verification already solve it?
  3. Do the subtasks need persistent collaboration?
  4. Is the value large enough to justify coordination cost?
  5. Can humans still understand and operate the system?

If the first two questions do not land clearly, the answer is often "stay simpler."

8. Quick comparison table

Dimension subagent agent team
purpose delegation, isolation, parallelism role separation, sustained collaboration
persistence shorter-lived longer-lived
state sharing can stay limited needs explicit design
orchestration lighter heavier
best fit independent subtasks ongoing collaborative work
main risk blurry delegation boundary exploding coordination cost

9. How this fits the rest of the series

The distinction becomes clearer when read against the earlier entries.

  • C2 argued that long work needs handoff before richer memory
  • C4 argued that memory ownership is about control of state
  • D1 treated role separation as a later-stage pattern
  • D4 showed public examples where orchestrator-worker and persistence matter differently

So the E4 conclusion is consistent with the rest of the series: multi-agent design is valuable only when responsibility boundaries need real separation.

10. Q&A 6: when is a single agent still better

More often than people expect.

  • the task is narrow
  • inputs and outputs are stable
  • validation is clear
  • role separation adds little

As D2 suggests, harness thickness should not grow by default.

11. Q&A 7: what subagents still require

Subagents are lighter than teams, but not free.

  • a clear delegated boundary
  • a lead integrator
  • stop rules for failed subwork
  • non-conflicting tool permissions

12. Q&A 8: what must be designed first in a team model

If a team is truly justified, define these before the team size.

Design item Why it matters
role definitions otherwise separation becomes decorative
output contracts handoff between roles needs stable formats
shared state memory and handoff rules must be explicit
approval boundaries execution and final approval should not blur
stop conditions loops need a termination rule

13. Final decision rule

Use this shortcut:

  • if a bounded subtask just needs to be handed off, use a subagent
  • if persistent collaboration and state coordination are central, consider a team

And above all:

  • complexity is not the default
  • do not build a team for a problem a single agent can already solve cleanly

14. What this appendix closes

E1 organized term boundaries. E2 organized evidence boundaries. E3 organized task-design boundaries. E4 organizes role-separation boundaries.

  • E1: term boundaries
  • E2: evidence boundaries
  • E3: task-design boundaries
  • E4: role-separation boundaries

The closing message is simple: multi-agent structure is justified only when the benefits of separation beat the cost of coordination.

References

  • docs/blog_series_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_์ด๊ด„_design.md
  • sources/260518_ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_15์žฅ_๋ธ”๋กœ๊ทธํ™œ์šฉ๋…ธํŠธ.md
  • drafts/blog/260519_Claudeํ•˜๋„ค์ŠคB02_CLAUDEmd_Skills_Hooks_Permissions_๋ธ”๋กœ๊ทธ.md
  • drafts/blog/260519_ํ•˜๋„ค์Šค์‹œ๋ฆฌ์ฆˆC02_์žฅ์‹œ๊ฐ„์—์ด์ „ํŠธ์šด์˜_๋ธ”๋กœ๊ทธ.md
  • drafts/blog/260519_ํ•˜๋„ค์Šค์‹œ๋ฆฌ์ฆˆD01_ํ•˜๋„ค์ŠคํŒจํ„ด12์„ _๋ธ”๋กœ๊ทธ.md
  • drafts/blog/260519_ํ•˜๋„ค์Šค์‹œ๋ฆฌ์ฆˆD04_์‹ค์ „์‚ฌ๋ก€๋กœ๋ฐฐ์šฐ๋Š”ํ•˜๋„ค์Šค์—”์ง€๋‹ˆ์–ด๋ง_๋ธ”๋กœ๊ทธ.md

This is Appendix Companion E4. This entry closes the appendix series.

Series overview: Harness Engineering Series Guide

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

"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 »

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 (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 »

"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 (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 »

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 »

"ML Foundations (6/9) — Neural Networks: From Perceptron to MLP"

์ด๋ฏธ์ง€
← 5/9 Regularization & Model… ๐Ÿ“š Series Index 7/9 Deep Learning Training → A neural network is logistic regression from Part 4 stacked into layers. Each layer passes through a new nonlinear activation, expanding the family of functions the network can represent. In this part we put how a neural network learns onto one page. Perceptrons, XOR, MLPs, backpropagation, activation functions, and the universal approximation theorem all meet here. 0. Learning Objectives Write the perceptron's learning rule and explain why it cannot solve XOR. Express an MLP's forward pass as one line of matrix products. Derive backpropagation by hand on a two-layer MLP. Compare the definitions, derivatives, and trade-offs of sigmoid, tanh, ReLU, and GELU. State the universal approximation theorem (Cybenko 1989 / Hornik 1991) precisely. Train an MLP end-to-end in PyTorch. 1. ํ•ต์‹ฌ ์š”์•ฝ Perceptron (1958): \(\hat{y} = \mathrm{step}(w^\top x + b)\). Converges only on linearly separabl...
Read more »