[EAAPL-AGT006] Reflexive Agent

Category: Agentic AI Sub-category: Quality Assurance Architecture Version: 1.1 Maturity: Emerging Tags: self-critique, reflection, quality-gate, generate-critique-revise, anti-loop, cost-control, output-quality Regulatory Relevance: EU AI Act (Art. 9, 15), ISO 42001 §8.4, NIST AI RMF (MEASURE 2.5)

1. Executive Summary

The Reflexive Agent Pattern defines an architecture in which an AI agent evaluates the quality of its own outputs through a structured generate-critique-revise cycle before returning results to the calling system. By adding an explicit self-evaluation step to the standard agent loop, organisations achieve measurable improvements in output quality — particularly for high-stakes knowledge work tasks like contract drafting, regulatory analysis, and clinical documentation — without requiring manual human review of every output.

For CIO/CTO audiences: this pattern is the AI equivalent of a professional practice quality review. A lawyer reviews their own memo before sending it; a radiologist performs a double-read on ambiguous scans. The Reflexive Agent embeds that review step into the automated workflow, catching errors and quality gaps before they reach users or downstream systems. The trade-off is cost: reflection requires additional LLM inference calls. This pattern defines the governance around when reflection is worth the cost, how to prevent reflection cycles from running indefinitely, and how to integrate reflection with human oversight. For high-stakes, low-volume tasks, the quality improvement easily justifies the cost. For high-volume, low-stakes tasks, reflection should be applied selectively based on confidence scoring.

2. Problem Statement

Business Problem

AI agents deployed for high-stakes knowledge work (legal drafting, medical documentation, financial analysis) produce outputs that are factually incorrect, structurally incomplete, or inconsistent with organisational standards at rates that are unacceptable for direct use without review. Manual review by human experts is the only existing quality gate, but it is expensive and creates the bottleneck that undermines the productivity value of automation.

Technical Problem

A standard agent loop generates outputs without any internal mechanism to evaluate their quality relative to the task objective. The model produces the most probable next token; it has no objective function that penalises factual errors, logical inconsistencies, or failure to meet specified quality criteria. Adding an external evaluation step after the loop completes catches errors too late — the full generation cost has already been incurred for an output that may require significant revision.

Symptoms of Absence

• Agent outputs for high-stakes tasks require expert human review of every output, negating the productivity benefit
• Quality is inconsistent and unpredictable — excellent outputs and poor outputs arrive with no distinguishing signal
• No feedback loop: the agent does not learn from its quality failures within or across tasks
• High escalation rate to human review even when outputs are clearly adequate

Cost of Inaction

• Quality Risk: Unreviewed poor-quality outputs from agents performing regulated tasks create compliance and liability exposure
• Operational: Expert review bottleneck grows with agent usage volume, offsetting scale benefits
• Competitive: Peers who implement reflection achieve demonstrably better output quality and can deploy agents in higher-stakes domains

3. Context

When to Apply

• Output quality has material business or compliance consequences (legal, medical, financial, regulatory)
• The task type has clear, articulable quality criteria that can be expressed in a critique prompt
• Task volume is moderate (the additional LLM cost per task is justified by quality improvement)
• The target quality improvement is measurable (a quality benchmark exists or can be created)
• Tasks where partial output correction is faster than full regeneration

When NOT to Apply

• High-volume, low-stakes tasks where reflection cost exceeds quality improvement value
• Tasks with no articulable quality criteria (purely subjective outputs)
• Real-time tasks with hard latency constraints incompatible with multi-pass generation
• Tasks where the initial output quality is already above the acceptance threshold (waste of compute)

Prerequisites

• EAAPL-AGT001 (Single Agent Pattern) baseline
• Defined quality rubric for the task type (criteria for the critique prompt)
• Quality threshold parameter (minimum acceptable quality score)
• Anti-loop detection (max revision iteration limit)
• Cost tracking per reflection cycle

Industry Applicability

4. Architecture Overview

The Reflexive Agent Pattern extends the standard agent loop (EAAPL-AGT001) by inserting a critique-revise sub-loop between the initial output generation and the final result delivery. The sub-loop has its own termination conditions and cost controls independent of the outer loop.

Why separate the critic from the generator? The same model that generates an output has a well-documented tendency to fail to critique its own errors — it is drawn toward confirming its own output rather than challenging it. Two strategies address this. First, the critique is prompted with an explicitly adversarial persona ("You are a strict expert reviewer. Identify all factual errors, logical gaps, and failures to meet the stated criteria"). Second, in higher-investment implementations, a separate model instance (or a different model entirely) performs the critique, reducing the correlation between generator and critic errors.

Generate Phase The initial generation follows the standard agent loop. The generate phase produces a candidate output — a document, analysis, code, or other artifact — and a confidence score (either model-produced or estimated from the output structure and completeness).

Confidence Gating Before entering the reflection sub-loop, a confidence gate evaluates whether reflection is needed. If the initial output's confidence score exceeds the configured "auto-accept threshold," the output is returned without reflection. This is the primary cost optimisation: for the majority of tasks where the initial output is clearly adequate, no additional inference calls are made. The threshold is tuned per task type based on observed quality distributions.

Critique Phase The Critique Engine receives the candidate output and the task objective (original instruction + quality rubric). It executes an LLM inference call with an adversarial reviewer persona. The critique prompt is carefully designed to produce structured output: a list of specific issues (each with a category: factual error / logical gap / missing requirement / style violation / inconsistency) and an overall quality score (0–100). The critique prompt is the most important engineering artefact in this pattern — vague critique prompts produce vague, unhelpful critique that does not guide revision.

Quality Gate The Quality Gate evaluates the critique output. If the quality score meets or exceeds the acceptance threshold and no critical issues are flagged, the output is accepted and returned. If issues are present, the Revision Engine is invoked.

Revision Phase The Revision Engine receives the original output, the original task instruction, and the structured critique. It invokes an LLM to produce a revised output that addresses the specific issues identified in the critique. The revision prompt is targeted: "Revise the following draft to address these specific issues: [critique issues]. Do not change content that was not flagged as an issue." This targeted revision approach is more efficient than full regeneration and preserves the valid portions of the initial output.

Anti-Loop Detection and Cost Control The reflection sub-loop enforces a hard maximum of N critique-revise cycles (default: 3). If the output has not reached the acceptance threshold after N cycles, the best output produced so far (highest quality score across all iterations) is returned with a reflection metadata flag indicating that the quality threshold was not reached. This prevents infinite reflection loops from running up unbounded inference costs. The total cost of all reflection cycles is tracked and reported; a per-task reflection cost ceiling can trigger early termination.

Reflection Memory Critique outputs from completed tasks are written to the agent's episodic memory store (EAAPL-AGT002) with the task type, initial quality score, final quality score, and the specific issues identified. The Memory Consolidation Engine processes these records to update the semantic memory with task-type-specific quality learnings. Over time, the generator's prompting is improved based on accumulated knowledge of the most common quality failures for each task type — reducing the number of reflection cycles needed and improving first-pass quality.

5. Architecture Diagram

6. Components

7. Data Flow

Full Reflection Cycle

Error Flow

8. Security Considerations

Prompt Injection in Critique

• The critique prompt injects the candidate output as content — if the candidate output contains injected instructions, the critique LLM could be manipulated
• Mitigation: the critique prompt wrapper clearly delineates the content being reviewed from the critic's instructions; content is wrapped in explicit delimiters (XML tags or similar); output validation on critique output before Quality Gate evaluation

OWASP LLM Top 10

9. Governance Considerations

Quality Rubric Governance

• Quality rubrics are owned by domain subject matter experts (legal team owns legal rubrics, clinical leads own clinical rubrics)
• Rubrics are versioned and change-managed; changes require impact assessment on existing task benchmarks
• Acceptance thresholds are set and reviewed by the domain owner, not by engineering

Model Risk Management

• Reflection quality scores are not objective ground truth; they are model judgments subject to model limitations
• Quality scores must be validated against human expert assessments on a held-out benchmark before being used as primary quality gatekeepers
• For highest-stakes tasks, model reflection quality scores are advisory only; human review remains the final gate

Governance Artefacts

10. Operational Considerations

SLOs

Monitoring

• Quality score distribution per task type: trending toward lower initial scores indicates prompt degradation
• Reflection cycle count distribution: bimodal (0 cycles or ≥2 cycles) may indicate confidence gate miscalibration
• Cost per reflection cycle per task type: anomaly detection for cost spikes

11. Cost Considerations

Cost Drivers

Optimisations

• Use a smaller, faster model for the critique step and the full model only for revision (model routing)
• Cache common critique patterns and their resolutions as procedural memories to reduce iteration count
• Tune confidence gate threshold upward (be more selective about what triggers reflection) if auto-accept rate is too low

Indicative Cost Range (per 1,000 tasks)

12. Trade-Off Analysis

Reflection Implementation Options

Architectural Tensions

13. Failure Modes

14. Regulatory Considerations

EU AI Act

• Art. 9 (Risk Management): reflection quality scores provide evidence of quality management for high-risk AI systems; must be preserved in the task audit log
• Art. 15 (Accuracy and Robustness): the reflection cycle directly implements the requirement for AI systems to remain accurate and robust; quality benchmark validation satisfies the measurement requirement

ISO 42001

• §8.4: The reflection mechanism and quality monitoring are part of the AI system's operational quality management lifecycle

NIST AI RMF

• MEASURE 2.5: The quality score time series and benchmark validation implement the AI performance measurement requirement

15. Reference Implementations

AWS

Azure

On-Premises

16. Related Patterns

17. Maturity Assessment

Overall Maturity: Emerging

18. Revision History