Unit 01.04: Deciding the smallest crew that could work

Crew designs tend to arrive with one agent per verb in the requirements. This is a way of testing that assumption before building it.

Three questions, two of which must be yes

Different tools, different permissions, different judgement. An agent clearing at least two of the three is doing something a merged agent could not; an agent clearing one is a task in disguise.

The code runs a five-agent proposal through the test.

PROPOSED = ["Researcher", "Summariser", "Fact checker", "Editor", "Publisher"]

QUESTIONS = [
    ("different tools?",       [True,  False, True,  False, True]),
    ("different permissions?", [False, False, False, False, True]),
    ("different judgement?",   [True,  True,  True,  True,  False]),
]

print(f"{'agent':14} " + " ".join(f"{q:22}" for q, _ in QUESTIONS) + "keep?")
for i, agent in enumerate(PROPOSED):
    answers = [values[i] for _, values in QUESTIONS]
    keep = sum(answers) >= 2
    cells = " ".join(f"{str(a):22}" for a in answers)
    print(f"{agent:14} {cells}{'yes' if keep else 'MERGE'}")

kept = sum(1 for i in range(len(PROPOSED))
           if sum(values[i] for _, values in QUESTIONS) >= 2)
print(f"\n{len(PROPOSED)} proposed, {kept} justified")

# Summariser and Editor each clear only one bar, so they are one agent with two
# tasks rather than two agents. Merging them removes a handoff, a context copy,
# and a place for the work to be dropped.

Summariser and Editor each clear only the judgement bar. They read text and produce text, with the same tools and the same permissions - so they are one agent with two tasks, and merging them removes a handoff, a context copy and a place for work to be dropped.

Publisher clears the permissions bar alone and is kept anyway, because write access is the one dimension where merging is actively dangerous. The rule is a heuristic; permissions override it.

The mistake this prevents

The mistake is treating the agent count as a measure of sophistication. It is a measure of coordination cost. Every agent is a context to fill, a handoff to get right, and a set of failures to debug, and the reviewer reading the run has to hold all of them.

Takeaway

Test each proposed agent against tools, permissions and judgement. Merge anything clearing fewer than two - except where write access is involved, where separation wins regardless.

Mark complete