Single Agent + Skills vs Multi-Agent Teams: When the Committee Loses

Most people reach for more agents when work gets hard.

Often the fix is the opposite: one agent, a skill library, and a green command - not a chatty committee.

This guide is for builders who watched multi-agent demos and want a setup that survives a real repo.

Kill the default approach

Default: spawn planner, coder, reviewer, and critic for every task. Hope debate produces quality.

What breaks first: token burn, conflicting edits, "done" with failing tests, and nobody owns the merge.

By the end you will have

1. When multi-agent helps vs when it wastes money
2. The Skill Library Compile magnet (turn roles into skills)
3. Isolation rules if you still need parallel research
4. Failure modes and a scorecard
5. A clear next action for one repo

The useful question

Not "how many agents?" - where does the work fail?

Research on multi-agent systems keeps finding the same pattern: coordination cost is real, and "teams" often fail to use the best expert in the room. Treat multi-agent as a special tool, not a default company structure.

When one agent + skills wins

Use a single agent when:

• One repo, one PR, one verification command
• The procedure is repeatable (skill-shaped)
• You need a consistent voice or coding style
• Cost and latency matter

Skill = muscle memory. The agent plans; the skill is the checked how-to.

Magnet: Skill Library Compile

Name: Skill Library Compile

Turn a multi-role fantasy into a single-agent skill shelf.

# From roles → skills
 
## Planner role becomes
skill: plan-diff
When: starting non-trivial work
Steps: state goal, list files, risks, done criteria
Verify: done criteria is testable
 
## Implementer role becomes
skill: implement-green
When: coding after plan
Steps: smallest diff, run check.sh, no drive-by refactors
Verify: check.sh exits 0
 
## Reviewer role becomes
skill: review-pr
When: before merge
Steps: security skim, test map, residual risks
Verify: risks listed; no secrets in diff

Load one skill at a time. Do not load the whole shelf into every prompt.

You should see: three skills used on real work twice each, with verify green, before you invent a fourth agent persona.

If you still need multi-agent

Only when all are true:

1. Task is parallelizable (independent sub-questions)
2. Workers get fresh context (no sibling chat)
3. One owner merges and verifies
4. Workers cannot publish, deploy, or spend

Parallel research is a good fit. Parallel "everyone edits main" is not.

Failure modes

One-page scorecard

Score 0-2 each: green command · skill count ≤10 · skills used twice · human gate on irreversible · no multi-agent public write.

0-4: demo mode. 5-7: reliable solo. 8-10: ready for optional research workers.

When not to use single-agent-only

• You need simultaneous web crawls across many sources (then isolated workers)
• A regulated review requires independent second pass
• You are measuring multi-agent systems on purpose

Path options

Bottom line

Prefer one brain and many skills over many brains and no verify.

Your next action: run Skill Library Compile on one repo and replace one multi-role prompt with three skills that each have a verify step.