Unit 09.03: When a second agent earns its cost

Multi-agent architectures are the most common over-engineering in this space. The question is not whether they work but whether one agent would have.

Relative cost against what you actually gain

Each additional agent adds context, a handoff to get wrong, and its own failure modes. The code prices four designs relative to a single prompt.

Read the note column alongside the cost.

DESIGNS = [
    ("one agent, one prompt",        1, 1.0,  "simplest, one place to debug"),
    ("one agent, two tool calls",    1, 1.4,  "still one context, still traceable"),
    ("two agents, handoff",          2, 2.6,  "separate contexts, a handoff to get wrong"),
    ("supervisor + three workers",   4, 5.1,  "routing bugs, plus three sets of failures"),
]

print(f"{'design':30} {'agents':>7} {'rel. cost':>10} note")
for design, agents, cost, note in DESIGNS:
    print(f"{design:30} {agents:>7} {cost:>10.1f} {note}")

print("""
A second agent earns its place when it has a genuinely different job, a
different tool set, or a different failure mode you want isolated -- not when
one prompt got long.

The honest test: could one agent with two tools do this? If yes, the multi-agent
version is costing 2.6x for an architecture diagram.
""")

A supervisor with three workers costs roughly five times a single agent and brings routing bugs plus three separate sets of failures to debug.

The honest test is in the last line: could one agent with two tools do this? Very often it could, and the multi-agent version is buying an architecture diagram at 2.6 times the price. A second agent earns its place when it has a genuinely different job, a different tool set, or a failure mode worth isolating - not when one prompt got long.

The mistake this prevents

The mistake is splitting an agent because its prompt became unwieldy. A long prompt is a prompt problem, and splitting it into two agents converts it into a distributed systems problem - with the long prompt now spread across two places that can disagree.

Takeaway

Add an agent when it has a different job, different tools, or a failure mode you want isolated. Prompt length is not one of those reasons.

Mark complete