The Ten Named SLR Agents
Every AI action inside a systematic review is attributed to a named agent, not a raw model identifier. This gives reviewers a stable mental model of who is doing what, and gives audit logs a human-readable signer rather than an opaque model string.
Why Named Agents Matter
- Reproducibility — a reviewer can cite "Screening Analyst v1 included this paper" regardless of which underlying model served the request.
- Trust — frontier models evolve quickly; the agent identity is stable.
- Audit clarity — the 21 CFR Part 11 chain records the agent, not the model, so chain verdicts survive model upgrades.
- UI consistency — every AI surface in the app (sidebar, generate buttons, activity feed) uses the same set of ten names.
The Ten Personas
Each SLR phase has exactly one named agent responsible for its AI capabilities:
| # | Phase | Agent | What it does |
|---|---|---|---|
| 1 | Protocol | Protocol Architect | Drafts protocol sections, PICO suggestions, eligibility criteria, search scoping |
| 2 | Search | Search Strategist | Query generation, MeSH / Emtree suggestions, database-syntax translation, dedup review |
| 3 | Screening | Screening Analyst | Include / exclude / uncertain recommendations with rationale and confidence |
| 4 | Extraction | Data Curator | Field-level extraction from PDFs, metadata reconciliation, provenance attribution |
| 5 | Risk of Bias | Quality Assessor | Domain-by-domain RoB 2 / ROBINS-I / QUADAS-2 judgments with rationale |
| 6 | Synthesis | Synthesis Engine | Narrative synthesis, heterogeneity commentary, subgroup/meta-regression framing |
| 7 | Report Prep | Compliance Officer | PRISMA / agency checklist compilation, SoF table framing, completeness review |
| 8 | Manuscript | Academic Writer | Section drafting, citation weaving, journal-style voice matching |
| 9 | Visuals | Visual Designer | Forest / funnel / flow / traffic-light rendering prompts and legends |
| 10 | Updates | Living Review Monitor | Re-search orchestration, new-paper triage, delta summary across updates |
How Attribution Appears
In the UI
Every AI-generated result panel shows the agent name prominently:
"Screening Analyst recommends: Include — PICO population matches, intervention matches dose range specified in protocol section 3.2."
Raw model identifiers (vendor slugs, model version strings, or anything resembling an API identifier) are never shown to the reviewer.
In the Audit Trail
Every audit event that originates from AI carries:
{
"aiAttribution": {
"agentId": "screening-analyst",
"agentLabel": "Screening Analyst",
"action": "recommend"
}
}
In the Activity Feed
The review's activity feed shows lines like:
10:42 — Search Strategist generated a PubMed query with MeSH expansion (2 rounds of self-critique)
11:03 — Screening Analyst marked 12 papers as "include" with mean confidence 0.86
Stable Across Model Changes
When the underlying model is upgraded, the agent identity is unchanged. The agent's capability profile (what tasks it accepts, what fallbacks exist) is code-owned and reviewed; the specific model that served a given request is an implementation detail of the routing layer.
This is especially important for regulatory submissions — an HTA / FDA auditor cares that "Screening Analyst" had a verifiable decision log, not which specific model weights ran that day.
Relationship to Multi-Agent Orchestration
The ten named agents are single-persona primitives. When a task benefits from multiple perspectives (e.g. manuscript section drafting with iterative critique), the orchestration layer composes multiple agents together — see the Agentic Orchestration article for details on debate, tree-of-thoughts, and judge-loop patterns.