Skip to content
Cortex

Synthesis sessions

Pressure-test a decision across 7+ agents in parallel.

What it is

A synthesis session asks one question across multiple CxO seats at the same time. Each agent answers from its own perspective, grounded in its usual topic files. Cortex then folds the answers into a single decision memo with positions, conflicts, and a recommendation.

Use it when a decision is genuinely cross-functional — pricing, hiring shape, vendor selection, market entry, security trade-offs, board-level positioning.

How to run one

  1. Phrase the question. One paragraph, decision-shaped, with the trade-off named. Bad: “Should we raise prices?”. Good: “Should we raise Growth prices by 20% in Q3, accepting 2-3pp churn risk in exchange for 14% MRR uplift?”.
  2. Pick the panel. Default to 7 — CEO + CFO + CTO + CMO + CISO + CRO + COO. Add specialists if relevant (GC for legal nuance, CDO for design risk).
  3. Pick the product. The session grounds in one product’s topic files.
  4. Run. Each agent answers in parallel; you can watch the streams interleave. Cortex generates a synthesis memo when all responses settle.

What you get back

The synthesis memo has four sections:

  • Positions — each agent’s 1-paragraph stance, with citations.
  • Conflicts — where the agents disagree and why.
  • Trade-offs — what each option costs.
  • Recommendation — Cortex’s synthesised call, marked as one input, not a mandate.

Cost

A 7-agent synthesis costs ~7× a single chat turn in tokens. Prompt caching helps — the topic-file bundle for a product is cached once and reused across all 7 calls. Plan for it: at Growth tier, two synthesis sessions a week comfortably fits.

Tips

  • Ask a sharper question. Vague questions get vague consensus. Trade-offs in the question force trade-offs in the answer.
  • Keep panels lean. 7 agents is usually right. 12 dilutes signal; 4 misses perspectives.
  • Don’t treat the recommendation as oracle. Read the conflict and trade-off sections — that’s where the value is.
  • Iterate. After the first synthesis, rerun with the refined question. Two passes are common.

Audit trail

Each agent call is recorded individually in anthropic_call_log. The synthesis memo records the full panel + question + per-agent reply IDs, so the entire decision is reconstructable months later.