OPENGAMP 5FDA/EMA GAIP (2026)ISO/IEC 42001

AI Agent Configuration Readiness

Treat an AI agent as a controlled configuration item and check whether its configuration is release-ready: model and version, system prompt, tools and permissions, knowledge sources, evaluation suite, approved Contexts of Use, and human oversight. Any missing mandatory element blocks approval deterministically — no approved Context of Use → not approved — because you cannot assure a configuration you cannot reconstruct. A scoping aid, not a validated assurance system.

OUTPUT

Release-ready / conditional / not-approved + missing elements

TIME

~8 min

Limitations — read before you rely on this

  • This is a scoping aid, not a validated system, and not a configuration-management system. Reproduce the configuration record in a controlled repository you own; a checklist result is not a configuration baseline.
  • It checks whether each element is declared present — not whether what you captured is correct, current, or adequate. A defined-but-wrong system prompt still reads as Defined here; the depth of each element is yours to assure.
  • Which elements are mandatory is SPEQ’s schema, not a regulator’s list. Your quality system may require more (or organize them differently); treat this as a starting frame to adapt, not a compliance checklist.
  • It assesses a single agent configuration at a point in time. A model or prompt change produces a new configuration to re-check — that change-impact step is a separate assessment.

WHAT THIS CALCULATES

Whether an AI agent’s configuration is reconstructable and therefore approvable. Treating the agent as a controlled configuration item — model and version, system prompt, tools and permissions, knowledge sources, evaluation suite, approved Contexts of Use, human oversight — it checks each element’s state and returns a deterministic verdict: release-ready, conditional, or not-approved. Any missing mandatory element blocks approval, because you cannot assure a configuration you cannot describe.

THE METHOD

verdict = not-approved if any mandatory element is undefined, else conditional if any applicable element is undefined, else release-ready; completeness = defined ÷ applicable
mandatory
an element required for a reconstructable configuration — identity, versions, system prompt, knowledge sources, eval suite, approved Context of Use, human oversight (and tool permissions once the agent has tools)
applicable
the elements in scope for this agent — every element except those marked genuinely not-applicable (N/A)
defined
the applicable elements that are actually specified and version-controlled
completeness
the fraction of applicable elements that are defined, in 0–100%
verdict
the deterministic outcome: release-ready, conditional (define the remaining optional elements), or not-approved (a mandatory element is missing)

Which elements are mandatory is SPEQ’s configuration schema expressing the epic’s determinism rules ("no approved Context of Use → not approved"), not a regulatory checklist. Tool permissions are mandatory only when the agent has tools. Monotonic: defining a missing element never worsens the verdict.

THE INPUTS, AND WHAT THEY MEAN

Element state (Defined / N/A / Missing)
For each configuration element, whether it is specified and version-controlled (Defined), legitimately not applicable to this agent (N/A), or absent (Missing). N/A removes an optional element from the completeness denominator; it cannot excuse a mandatory one.
Whether the agent has tools
Set by marking the Tools element Defined. It flips tool-permissions from optional to mandatory, because an agent that can act must have its permissions and their limits defined — expanding a permission later requires a change-impact assessment.
[ AGENT CONFIG · AI ASSURANCE ]

Is this agent’s configuration reconstructable and approvable?

Treat an AI agent as a controlled configuration item. Mark each element Defined, N/A, or Missing, and SPEQ returns whether the configuration is release-ready. Any missing mandatory element blocks approval deterministically — you cannot assure a configuration you cannot reconstruct, and no approved Context of Use means not approved.

Agent identityMANDATORY
A named, uniquely identified agent — you can say exactly which agent this configuration is.
Agent versionMANDATORY
The exact qualified version of the agent as a whole, so a run can be tied to a configuration.
Model + providerMANDATORY
The underlying model and its provider are named and replaceable.
Model versionMANDATORY
The exact model version the agent runs on — a provider update is a change that can trigger reassessment.
System promptMANDATORY
The governing system prompt is captured and version-controlled, not edited in place untracked.
Prompt templates
Any task or prompt templates the agent uses are captured with the configuration.
Tools
The tools the agent can call are enumerated — you know what it can do.
Tool permissions
What each tool is permitted to do, and its limits, are defined. Expanding a permission requires a change-impact assessment. Mandatory once the agent has tools.
Knowledge sourcesMANDATORY
The sources grounding the agent’s regulatory knowledge are identified, authoritative, and current.
Retrieval configuration
Chunking, retrieval, and vector-store settings are recorded so retrieval behaviour is reconstructable.
Embedding model
The embedding model and version behind retrieval are named — a change re-embeds the corpus.
Guardrails & policies
Input/output guardrails and operating policies are defined and version-controlled.
Evaluation suiteMANDATORY
A qualifying evaluation suite is bound to the agent and its Contexts of Use (see the evaluation foundation).
Approved Contexts of UseMANDATORY
The specific Contexts of Use the agent is approved for are recorded. No approved Context of Use → not approved.
Human oversightMANDATORY
The review/approval level for the agent’s outputs is defined and meets the risk-based floor.
Monitoring
How continued performance is detected in production is defined — the signals that raise an assurance event.
CONFIGURATION VERDICT
NOT APPROVED
0/16 applicable elements defined · 0% complete.
MANDATORY — BLOCKS APPROVAL
  • Agent identity
  • Agent version
  • Model + provider
  • Model version
  • System prompt
  • Knowledge sources
  • Evaluation suite
  • Approved Contexts of Use
  • Human oversight
OPTIONAL — COMPLETE THE RECORD
  • Prompt templates
  • Tools
  • Tool permissions
  • Retrieval configuration
  • Embedding model
  • Guardrails & policies
  • Monitoring
WHY

9 mandatory configuration elements are undefined (e.g. identity) — the configuration is not reconstructable, so it cannot be approved. Determinism, not judgment.

PROFESSIONAL EXPORT

HOW TO READ THE OUTPUT

  • Not-approved is a statement about reconstructability, not quality. A brilliant agent whose system prompt or knowledge sources are not captured cannot be assured, because a future reviewer cannot reproduce what it was — so the gate refuses to approve it.
  • Conditional means the configuration is safe to reason about but not yet a complete record: every mandatory element is defined, but optional elements (retrieval config, guardrails, monitoring) are still worth capturing before the configuration is considered final.
  • Use N/A honestly. A retrieval-augmented agent cannot mark knowledge sources N/A; a non-RAG agent legitimately can mark retrieval config and embedding model N/A. N/A narrows the denominator, so misusing it inflates completeness against you at audit.
  • The verdict is the entry gate to assurance, not the whole of it. A release-ready configuration is one you can now assess for risk, evaluate, and monitor — it does not by itself mean the agent is validated.

WORKED EXAMPLE

A RAG regulatory copilot with tools. Everything is captured except the approved Contexts of Use, which have not been formally recorded yet.

Model, version, system prompt
Defined
Tools + tool permissions
Defined
Knowledge sources, eval suite, oversight
Defined
Approved Contexts of Use
Missing

RESULT

Verdict: NOT-APPROVED — a mandatory element (approved Contexts of Use) is undefined, ~93% complete

The configuration is almost complete and still not approvable: without a recorded approved Context of Use there is no answer to "approved for what?", so the deterministic rule blocks it regardless of how much else is in place. The fix is specific — record the Contexts of Use the copilot is approved for — after which the same configuration becomes release-ready.

REGULATORY BASIS

FDA/EMA — Guiding Principles of Good AI Practice in Drug Development (Jan 2026)
Frames the lifecycle and configuration control this registry supports; see SPEQ’s Good AI Practice page.
ISPE GAMP 5 (2nd ed.) — A Risk-Based Approach to Compliant GxP Computerized Systems
Supplies the configuration-management and version-control expectations that make a system’s behaviour reconstructable.
ISO/IEC 42001:2023 — AI management systems
The management-system requirement to control AI system configurations and changes across the lifecycle.
PROFESSIONAL · WORKED SCENARIOS · SPEQ SYNTHESIS

See this tool applied to real cases

CHECKING ACCESS

Checking your Professional access…