Research surface

Signal pipeline

A workflow typically chains six tasks. Each is configurable — prompts, models, thresholds, parallelism — and can be re-ordered or omitted. Together they take a briefing all the way through to a report.

1 — Generate

One or more AI subagents search the web (using model-native search) and synthesize an initial pool of candidate signals against your briefing and categories. Subagents can use different models so you get a diverse first pass.

2 — Merge

A single LLM pass consolidates duplicates and near-duplicates, picking the strongest version and unioning its supporting sources. Originals are kept (soft-merged) so you can audit what was combined.

3 — Source and verify

Per signal, an agent re-checks existing sources, hunts for new ones, judges each as supports, contradicts, unrelated, or unreachable, then commits to a signal-level verdict — verified or rejected — with a confidence level and rationale.

4 — Assess

One task per metric. The agent sees every active signal at once and scores them comparatively against that metric using the scale you defined. Multi-model assessments reconcile via median or mode.

5 — Image

For each signal, the agent renders a prompt template against the signal’s content, calls an image model, and attaches the result. Signals that already have an image are skipped unless you force regeneration.

6 — Report

The agent reads all active signals with sources, grounding, scores, and tags and writes a report shaped by the report type’s prompt. Each run adds a new report — nothing is overwritten.

Related