· 5 min read
Every agent team pattern ends with some form of synthesis — the step where individual agent outputs get combined into a final deliverable. This step is where agent teams either produce something genuinely better than a single agent could, or produce an expensive mess.
The difference is almost always the synthesis prompt.
Synthesis is not concatenation. Slapping three agent outputs together with section headers is not synthesis. It's a document with three sections.
Synthesis is not summarization. Shortening each agent's output and combining the summaries loses the detail that makes multi-agent analysis valuable.
Real synthesis does something none of the individual agents could do alone: it finds connections across outputs, resolves contradictions, identifies emergent patterns, and produces insights that only become visible when you look at all the pieces together.
A competitive intelligence team's synthesizer doesn't just list what each agent found. It connects the pricing analyst's findings to the feature mapper's findings: "Competitor X dropped their price by 20% the same quarter they removed three features — they're repositioning downmarket, which opens an opportunity in the enterprise segment."
That insight exists in neither agent's output individually. It only emerges in synthesis.
Agent outputs will contradict each other. One agent might assess a market as growing; another might flag declining customer interest. A strong synthesis prompt doesn't ignore this — it requires the synthesizer to address it.
Include in your prompt: "Identify any contradictions between agent outputs. For each contradiction, evaluate the evidence on both sides and provide your assessment of which position is better supported, with reasoning."
The most valuable insights often span multiple agents' domains. A theme that appears in the market research, the competitive analysis, and the customer data simultaneously is a stronger signal than any single finding.
Include in your prompt: "Identify themes or patterns that appear across multiple agent outputs. Rank these cross-cutting themes by the number of agents whose findings support them and the strength of the supporting evidence."
Not everything matters equally. A synthesis that lists 30 findings with equal weight is almost as useless as no synthesis at all. The synthesizer must rank.
Include in your prompt: "Prioritize the top 5-7 findings by business impact. For each, explain why it ranks where it does and what action it implies."
Insights without actions are academic exercises. The synthesis must bridge from "what we found" to "what we should do."
Include in your prompt: "For each top finding, provide a specific, actionable recommendation. Each recommendation should include who should act, what they should do, and what the expected outcome is."
Sometimes the most important insight is what's missing. The synthesizer should identify questions that the agents couldn't answer, data that would strengthen the analysis, and areas where confidence is low.
Include in your prompt: "Identify any gaps in the analysis — questions that remain unanswered, areas where evidence is thin, or topics that none of the agents adequately addressed. Recommend follow-up actions to close these gaps."
Weak:
Combine the outputs from all agents into a final report. Summarize the key findings from each agent.
This produces a book report. Each agent gets its own section with a summary. No connections, no contradictions resolved, no new insights.
Strong:
You are the Strategic Synthesizer. You will receive outputs from three specialist agents: a Market Researcher, a Competitive Analyst, and a Customer Insights Analyst.
Your job is NOT to summarize their work. Your job is to find what none of them could find alone.
Specifically:
- Identify contradictions between agent outputs and resolve them with evidence-based reasoning
- Surface cross-cutting themes that appear in 2+ agent outputs — these are your highest-confidence findings
- Prioritize the top 5 findings by business impact, with a one-sentence justification for each ranking
- For each top finding, provide a specific recommendation: who should act, what action to take, expected outcome
- Flag 2-3 gaps where the analysis is incomplete and recommend follow-up actions
Structure your output as: Executive Summary (200 words), Key Findings (ranked), Recommendations (actionable), and Open Questions (with suggested next steps).
The difference is night and day. The strong prompt produces a deliverable that a leadership team can act on. The weak prompt produces a document that needs another round of human synthesis to be useful.
Contradictions aren't bugs — they're signals. When agents disagree, it usually means the underlying question is genuinely ambiguous. Here's how to handle it:
Don't force resolution. Sometimes the right answer is "the data supports both interpretations." A synthesis prompt that forces the synthesizer to pick one side loses valuable nuance.
Require evidence evaluation. When agents contradict each other, the synthesizer should assess the quality and recency of each agent's evidence, not just pick the more optimistic take.
Present conditional conclusions. "If the market is growing (as Agent A suggests), then Strategy X is recommended. If demand is softening (as Agent B suggests), then Strategy Y is the safer path." This gives decision-makers a framework instead of a guess.
Adapt this to your use case:
You are the [Role Name]. You will receive outputs from [N] specialist agents: [list agents and their roles].
Your job is to produce a synthesis that creates insights none of the individual agents could produce alone.
Contradictions: Identify where agent outputs conflict. Evaluate the evidence and provide your assessment.
Cross-cutting themes: Surface patterns that appear across multiple agent outputs. Rank by confidence.
Priority findings: Select the top [5-7] findings by [impact metric]. Justify each ranking.
Recommendations: For each finding, provide an actionable recommendation with owner, action, and expected outcome.
Gaps: Flag [2-3] areas where the analysis is incomplete and suggest follow-up.
Output format: [Specify sections, word counts, structure]
Fill in the brackets, adjust the numbers, and you have a synthesis prompt that produces genuinely useful output.