orchestration-patternslisted

# Orchestration patterns: research that fact-checks itself When you fan out subagents to research something, the danger isn't too little output — it's confident, plausible, **wrong** output that reads as authoritative because a fleet of agents produced it. These two templates exist to make a fan-out fact-check itself, so what survives to your synthesis is the part that withstood an adversary. Both templates are in `templates/`. Both are domain-agnostic — you drive them with `args`, not by editing the script. Fill in a worked example by reading `references/worked-example-observability.md`. ## The shared discipline (why these work) Three invariants hold across **both** templates. They are the reusable idea; the two shapes below are just two ways to spend them. 1. **Frozen ground-truth context, injected into every agent.** One `context` string carries the facts you already trust (hard constraints, confirmed environment, known-good values). Every agent — researcher and verifier alike — receives it verbatim, so no agent reasons from unanchored assumptions and they don't contradict each other on settled facts. In the templates this is `CTX`, marked `FROZEN-CONTEXT`. 2. **Per-stage structured output.** Every stage forces a JSON `schema`, so the next stage consumes a validated shape instead of parsing prose. This is what makes the *harvest* step possible — you can only pull "risky claims" out of a research result if the result is guaranteed to have a `ri