Same problem, N agents, different starting conditions. Then one agent assembles the best parts of each — not a vote, not a summary, a deliberate merge.
Why this, for you: orchestrator-worker splits one task into many subtasks. Fan-out
synthesis runs one task N times and combines the attempts. It's the right tool for high-stakes design and
architecture decisions — but the synthesis agent is the single highest-risk component, and getting it wrong makes
the output worse than your best single attempt.
A single LLM call samples the solution distribution once. N independent calls sample it N times,
covering more of the space. Synthesis exploits that variance instead of averaging it away.
1 The three-step shape
Fan-out — spawn N agents with identical instructions but independent contexts; each produces a distinct solution.
Synthesis — a synthesis agent critiques all N outputs, scores them against defined criteria, and assembles a merged solution from the strongest parts.
Validation — pass the merged output through a committee review loop before accepting it.
Unlike majority voting, which picks the most popular answer, synthesis
combines complementary strengths deliberately — Agent 1's naming, Agent 2's versioning, Agent 3's
error handling — into a composite no single agent reached.
2 Diversity is the whole engine
The mechanism is ensemble variance reduction applied to generative output — combining diverse weak learners
outperforms any individual one. But the key condition is genuine diversity: if the agents converge on
the same approach, there's nothing for synthesis to exploit. Identical instructions do not guarantee distinct
outputs, so you engineer the spread:
Vary temperature across agent instances
Vary seed context — a different starting reference for each
Vary system-prompt emphasis — one optimizes for brevity, another robustness, a third edge-case coverage
Conformity bias collapses the diversity you paid for
Agents given the same prompt can converge rather than explore independently — inter-agent misalignment is one of
three categories in the multi-agent failure taxonomy. Constrained solution spaces amplify the convergence. Running
more agents does not guarantee the diversity fan-out depends on; you have to force the spread.
3 How many, and the riskiest part
N parallel attempts cost N× compute, and the returns diminish fast. Best-of-N research shows quality gains compress
while compute grows linearly — making N = 3–5 the efficient range. N=10 rarely justifies 10× the cost
over N=3.
The synthesis agent is the highest-risk component. If it can't judge which elements
are strongest, the merge step introduces errors rather than removing them — and the result can be worse than
the best individual attempt. Synthesis is deliberate assembly with justification, not summarization.
Two more guards: passing all N outputs to one synthesizer can overflow context, and when the merged output feeds a
downstream agent as authoritative, synthesis errors compound instead of self-correcting. That's why the
pattern chains into a committee review — fan-out generates diversity; review validates the merge
before anyone accepts it. Worthwhile for high-stakes or creative work; overkill for routine, well-defined tasks where
a single attempt suffices.
↪ Your win: exploit variance, then validate the merge
Run one task N times with independent contexts — distinct from splitting one task into subtasks.
Engineer diversity — vary temperature, seed context, or prompt emphasis; convergence kills the gain.
Synthesize, don't vote or summarize — deliberately assemble the strongest parts, with justification.
Stay in the N = 3–5 range — gains compress while compute grows linearly.
Chain into committee review — the synthesizer is the highest-risk step; validate before accepting.
Retrieval practice — recall, don't peek
Question 1Synthesis differs from majority voting because it…
Question 2Fan-out synthesis only works when the N agents produce…
Question 3The efficient range for N, where gains stop compressing, is about…
Question 4The single highest-risk component of the pattern is the…
Question 5 · spaced recall from Lesson 2In orchestrator-worker, the orchestrator's final job is to…
Ask me anything. Want a fan-out setup for a real design decision — three agents with distinct
emphases plus a synthesis rubric — or how to wire the committee-review gate after it? That closes Part 1. Next:
Forked vs Fresh Subagents — the first coordination contract.