E1 · The matrix in the example
10 minClone the practice project, save decide.py and tarefas.txt, and run it.
Done when: python3 decide.py tarefas.txt | grep -c "^decisão:" → 4, and the market research comes out as decisão: híbrido.
Lesson 1 · 15 min
When NOT to use multi-agent
Every extra agent reads the context again, spends its own quota, and has to be coordinated. For most code tasks, a single agent, or subagents that return little, gets you more. In this lesson you decide before opening the second agent, using a matrix of signals that runs as a script.
Start the lesson
In this lesson you will
- 1What a team costs: the published numbers and what they measure
- 2The 6 sequence signals and the 7 parallel ones
- 3The 2-point difference rule and the hybrid
- 4The matrix as a script, with your own tasks
Goal: decide between a single agent, sequence, parallel, or hybrid using the signal matrix, before opening the second agent.
01Before the second agent
Click each step. The matrix comes from the arquiteto-de-execucao kit (agenteexecuta).
Click each block in the diagram to see what it does.
The matrix is a starting point, not a measurement. It prevents the reflex of opening agents just because you can.
02The cost that doesn't show up in the request
What is published, with dates, and what each number measures.
1. Tokens
In Anthropic's post about its research system (06/13/2025): agents use about 4× the tokens of a chat, and multi-agent systems about 15×. The data is theirs, from a research product.
The Claude Code agent teams doc says the same without a number: a team uses significantly more tokens than a single session, and the spend grows with each active teammate.
2. Gain
In the same post, an Opus 4 lead with Sonnet 4 subagents beat Opus 4 alone by 90.2% on their internal research evaluation, on breadth questions (many independent directions).
The post itself warns: most coding tasks have fewer parallelizable parts than research, and agents still coordinate poorly in real time. No measurement from this course shows that gain carrying over to your code.
3. Context
The arquiteto-de-execucao matrix shows the math (illustrative example, not measured): three agents that read 50k of context and work 20k each add up to 210k; one agent that reuses the context does the same with 110k.
The savings come from not repeating the setup. Parallel pays off when each task has small context.
4. Limits
No subagent without an explicit task, input, and output; at most one level of delegation, unless clearly necessary; two attempts per item before rethinking; don't reprocess an approved item.
Mechanical work (converting, rendering, running tests) goes to a script, not an agent.
03The matrix as a script
Write each task's signals in a file; the script adds them up and decides. No model involved: it runs for free.
Terminal · course training project (once)
git clone https://github.com/inematds/inema-agent-runtime.git ~/treino-multi cd ~/treino-multi && git log -1 --format='%h %cs' # testado em 5cb431b 2026-10-05
decide.py · save it in ~/treino-multi/
# decide.py — decision matrix of the arquiteto-de-execucao (inematds/agenteexecuta), in code
# Usage: python3 decide.py tarefas.txt
# tarefas.txt: one block per task; "# nome" opens the block; then, one code per line:
# S1..S6 = sequence signals, P1..P7 = parallel signals,
# simples = small task (the matrix does not apply), erro-caro = high cost of error
import sys
SEQ = {"S1": "depende da saída anterior", "S2": "contexto compartilhado grande",
"S3": "decisões precisam ficar consistentes", "S4": "erro inicial contamina várias tarefas",
"S5": "juntar versões divergentes sai caro", "S6": "uma etapa ensina a seguinte"}
PAR = {"P1": "tarefas independentes", "P2": "entradas congeladas",
"P3": "cada saída se valida sozinha", "P4": "trabalho mecânico ou determinístico",
"P5": "contexto pequeno por tarefa", "P6": "urgência alta", "P7": "há máquina sobrando"}
def decide(nome, cods):
print(f"# {nome}")
for c in cods:
if c not in SEQ and c not in PAR and c not in ("simples", "erro-caro"):
print(f"aviso: código desconhecido '{c}' (ignorado)")
if "simples" in cods:
print("decisão: direto (um agente só, sem matriz)\n")
return
s = sum(c in SEQ for c in cods)
p = sum(c in PAR for c in cods)
if abs(s - p) >= 2:
d = "sequencial" if s > p else "paralelo"
else:
d = "híbrido"
print(f"sequência: {s} · paralelo: {p} · diferença: {abs(s - p)}")
print(f"decisão: {d}")
if "erro-caro" in cods:
print("custo de erro alto: planejamento e revisão centralizados")
print()
nome, cods = None, []
for linha in open(sys.argv[1], encoding="utf-8"):
linha = linha.strip()
if linha.startswith("#"):
if nome:
decide(nome, cods)
nome, cods = linha.lstrip("# ").strip(), []
elif linha:
cods.append(linha.split()[0])
if nome:
decide(nome, cods)tarefas.txt · four example tasks
# refactor the login module S1 each step uses the previous one S2 the whole module has to be in your head S3 same names and contracts everywhere S4 an error in the contract breaks the rest erro-caro # convert 500 .md files to .html P1 one file doesn't depend on another P2 the converter and the model are locked P3 each .html checks itself P4 it's mechanical P5 each file is small # market research across 4 segments S3 the questions have to be the same S6 the first segment tunes the questions P1 independent segments P5 little context per segment # summarize a text in 5 lines simples
Terminal · run
cd ~/treino-multi python3 decide.py tarefas.txt; echo "saída: $?"
Example from the repository
# refatorar o módulo de login sequência: 4 · paralelo: 0 · diferença: 4 decisão: sequencial custo de erro alto: planejamento e revisão centralizados # converter 500 arquivos .md para .html sequência: 0 · paralelo: 5 · diferença: 5 decisão: paralelo # pesquisa de mercado em 4 segmentos sequência: 2 · paralelo: 2 · diferença: 0 decisão: híbrido # resumir um texto em 5 linhas decisão: direto (um agente só, sem matriz) saída: 0
Converting 500 files is mechanical parallel work: a script with several workers and a sample check at the center (that's lesson 3). The market research came out tied: the first segment runs alone, the other three in parallel.
A lowercase code doesn't count. The script warns you instead of quietly adding it up, and the decision changes. Captured the same day:
# migrar o banco aviso: código desconhecido 's1' (ignorado) sequência: 1 · paralelo: 1 · diferença: 0 decisão: híbrido
With the correct S1, it would be sequence 2 × parallel 1, still hybrid, but for a reason you chose. Check grep -c "^aviso:" decisao.md → 0.
04Practice
First the example, then three tasks of your own.
E2 · Three tasks of your own
20 minWrite minhas-tarefas.txt with three real tasks from your week (one you've thought about splitting between agents). Mark only the signals that are true today.
Done when: python3 decide.py minhas-tarefas.txt > decisao.md; grep -c "^decisão:" decisao.md → 3 and grep -c "^aviso:" decisao.md → 0.
05Lesson quiz
Question 1 of 4
According to Anthropic's post (Jun/2025), how many tokens does a multi-agent system use compared to a chat?skill
Correct
About 15× (and single agents, about 4×), in the data from their research system. Splitting the work doesn't split the context: each agent reads its own.Not this one
About 15× (and single agents, about 4×), in the data from their research system. Splitting the work doesn't split the context: each agent reads its own.Review: The cost that doesn't show up in the requestQuestion 2 of 4
Sequence 3, parallel 2. What does the matrix say?skill
Correct
The difference is 1, less than 2: hybrid. The common mistake is picking the bigger side for any difference at all.Not this one
The difference is 1, less than 2: hybrid. The common mistake is picking the bigger side for any difference at all.Review: Before the second agentQuestion 3 of 4
A migration with a high cost of error scored parallel. What changes?
Correct
A costly error centralizes the plan and the review, whatever the score. Running the independent items can still be parallel.Not this one
A costly error centralizes the plan and the review, whatever the score. Running the independent items can still be parallel.Review: Before the second agentQuestion 4 of 4
Why doesn't the research system's 90.2% gain apply directly to your code?
Correct
The post itself makes this caveat: the gain shows up in questions with many independent directions. In code, dependencies weigh more.Not this one
The post itself makes this caveat: the gain shows up in questions with many independent directions. In code, dependencies weigh more.Review: The cost that doesn't show up in the requestQuiz completed
You got {a} of {n} right. Get them all right and the lesson is marked as completed.Choose one option. The explanation appears after your attempt.
06In three sentences
A team costsEach agent rereads the context and spends its own quota; published: about 15× the tokens of a chat.
Matrix firstSigns of sequence × parallel; a difference of 2 decides, less than that is hybrid; costly errors centralize.
Simple means directA small task doesn't go through the matrix, and mechanical work goes to a script.