Conference Pipeline Orchestration

This skill is the operating discipline for an orchestrator that runs a staged pipeline of specialist agents — for a conference scheduler that is ingest → cluster → elicit → schedule, but the discipline is general. Its single organizing idea is explicit uncertainty management at every layer: each stage produces calibrated confidence, not commitments, and the orchestrator's whole job is to move that uncertainty forward honestly, integrate it, decide when it is small enough to act on, and refuse to let its own safety machinery be optimized away.

The discipline is drawn from how collectives of interacting agents actually fail. The leverage is not in the loudest stage; it is in the shape of the communication between stages and in a few feedback loops on safety-critical asymmetries. So the rules below are mostly about the edges of the pipeline (how stages talk) and about a small set of invariants (what the orchestrator must never let drift), not about making any single stage cleverer.

Common Patterns

Pattern 1: Confidence propagates — every stage emits it, the next stage reads it

Each stage attaches calibrated confidence to its output and the orchestrator carries it forward: extraction tags low-confidence fields, clustering exposes multi-membership and an outlier bucket, elicitation tracks a preference region (uncertainty bands), scheduling exposes a weighted objective and surfaced conflicts. The orchestrator does not collapse this into a single "done" — it reads the confidence at each gate and lets it govern whether to advance, reconcile, or escalate. A pipeline that throws away per-stage confidence cannot make an honest final recommendation.

Orchestrator action: at each gate, read the artifact's meta/status block (e.g., confidence_rollup, region_tightness, unresolved_conflicts) and branch on it.

Pattern 2: Structured communication contract — files and status fields, never free text

Agents communicate through artifacts with defined schemas, not through prose summaries. The orchestrator spawns a worker with: the input paths to read, the parameters (ids, output dir), an explicit output path to write, and a return contract ("return just the path"). The worker writes the schema-shaped artifact and returns the path; the orchestrator verifies the artifact's status fields before advancing. This is what makes the pipeline auditable and resumable — the canonical state lives in files, and any agent (or human) can read exactly what was passed and produced. Free-text handoffs are where pipelines silently corrupt.

Orchestrator action: never hand a worker a paraphrase of upstream output; hand it the path. Never trust a worker's prose claim of success; read its artifact's status.