Agent operating model

What "an agent" means here

An agent is an AI assistant given a narrow, written brief and a fixed set of approved sources — for example one return's published specification and the local data dictionary. It answers questions and drafts text strictly within that brief, always citing where its answer came from, and stops and escalates when something falls outside its sources. Think of it as a well-briefed, source-bound assistant for one task — not a general chatbot, and never the decision-maker.

Core principles

Pairing agents with approved sources

Reporting specification agents — for the Mental Health Services Data Set (MHSDS), Community Services Data Set (CSDS), Patient Level Contract Monitoring (PLCM) and Aggregate Data Collections (ADC) — should be paired with:

• Downloaded public specification documents (version-dated)
• Local data dictionary (docs/local-dictionary-placeholder.md in demo; MHSDS worked example uses docs/synthetic-mhsds-local-dictionary.md)
• MHSDS public source pack index (public-data/mhsds-source-pack-register.csv) for the SME worked example
• Local SOP or process notes (IG-approved; not published here)
• Source restrictions and citation requirements in the agent rule file
• Named human sign-off points before operational use

Example questions SME agents should support

Each answer must cite the source section so the user can check at source. If the answer is not in approved documents, the agent must say so and escalate.

Worked example: asking an MHSDS SME agent about an unusual figure

In plain English: when a number on a report suddenly jumps, the first job is to find out why — is it a genuine change in activity, or just a change in how things were recorded or counted? "Tracing a figure back to source" means following that number step by step, from the original record a clinician entered through to the final total on the report, checking each stage. The technical chain below is kept deliberately, because that traceability is the whole point — but the takeaway is simple: the agent helps you ask better questions, and a person still decides.

This example shows how a source-bound SME agent could support a Business & Performance user investigating an unusual movement in a synthetic MHSDS-derived metric. The agent traces the figure back to source forms, extract logic, fact tables and mapping rules, separates data-quality from operational explanations, and offers draft pack wording — sign-off is discussed only when the user asks whether the figure is ready for the pack.

What this demonstrates

Practical data lineage (RiO contact → staging extract → fact table → team mapping → count), movement calculation (+370 contacts, ~29.8%), data-quality vs operational hypotheses, numbered investigation checks, and draft directorate pack wording pending human review.

User:

MHSDS Expert Agent:

User:

MHSDS Expert Agent:

Why this matters: The value of an SME agent is tracing a reported number back to RiO forms, extract tables, fact tables and mapping rules — and knowing what data-quality and operational checks to run. It helps the user ask better questions and draft proportionate caveats; it does not replace human validation or sign-off.

Supporting files:

• MHSDS Expert Agent rule
• Synthetic local dictionary
• Read full conversation transcript →
• MHSDS source pack register
• Synthetic aggregate demo data

Worked example: Report Analysis Agent reviewing a draft Talking Therapies brief

In plain English: before a draft report is shared, someone needs to check that it answers the basic performance questions — what the figure is, what it should be (where a standard exists), how it is trending, what judgement is fair, and what a human must still verify. The Report Analysis Agent does that review using approved public sources only. It cites evidence, suggests draft fixes, and never signs the report off.

This example uses a synthetic pre-fix draft (deliberate errors) based on issues found during the June 2026 public-data report improvement pass. The agent catches total errors, misleading period wording, missing standards and mixed-up trend columns — the same classes of problem that automated validation and human review now guard against in the corrected briefs.

What this demonstrates

Eleven-step review framework: period/total/standard/trend-column/caveat/bottom-line checks; CSV traceability; revised wording marked for human review; refusal when asked to approve publication. Cross-references 04_validate_public_reports.R where mechanical checks apply.

User:

Report Analysis Agent:

User:

Report Analysis Agent:

Why this matters: This agent operationalises the AI assurance checklist in practice — the same issue classes that drove the June 2026 public-data improvement pass. See the corrected Talking Therapies brief for the post-fix outcome. For source-scope patterns (A&E without ED), see the urgent diagnostics brief.

Supporting files:

• Report Analysis and Improvement Agent rule
• Flawed draft input (v0.3)
• Read full conversation transcript →
• Corrected brief (after improvement pass)
• Demo CSV source
• Post-render validation script

Workflow (conceptual)

Reporting workflow

For performance reporting and mandatory returns: interpret data, check definitions, quality-assure (QA) drafts and prepare narratives — each step produces drafts or findings for human review, not autonomous outputs.

Admin and delivery workflow

For delivery work: coordinate project actions, reconcile backlogs and close down branches before merge or handover — still subject to IG/Safety rules. Neither workflow implies autonomous changes to the code repository (Git), the project-tracking tools (such as Azure DevOps) or any external system. See agent-rules/README.md.

A. Reporting specification SME agents

B. Performance and reporting workflow agents

C. Admin and delivery agents

D. Information governance