// PROJECT · 2026

Self-Improving Agentic
Solution Designer

Interview-driven multi-agent system using 17 Claude Sonnet subagents to generate fit-for-purpose enterprise architecture artefacts via intent-signalled selection. Features bi-temporal memory consolidation, skill graduation gates, adversarial critique loops, and a quality-gated PlantUML→draw.io pipeline producing 18 enterprise-grade outputs.

17 SubagentsMulti-AgentADRsC4Threat ModelsNFRsPlantUMLdraw.ioBi-temporal Memory

Architecture

┌─────────────────────────────────────────────────────┐
  │              INTERVIEW ORCHESTRATOR                  │
  │         (intent-signal extraction layer)             │
  └──────────────┬──────────────────────────────────────┘
                 │ intent signals
    ┌────────────▼────────────┐
    │   SELECTOR ROUTER       │  ← chooses 1..N subagents
    └────────────┬────────────┘
                 │
   ┌─────────────┼──────────────┐
   ▼             ▼              ▼
[ADR]  [C4 + PlantUML]  [Threat Model]  …14 more
   │             │              │
   └─────────────┼──────────────┘
                 │
    ┌────────────▼────────────┐
    │   ADVERSARIAL CRITIC    │  ← adversarial critique loop
    └────────────┬────────────┘
                 │ passes quality gate?
    ┌────────────▼────────────┐
    │   BI-TEMPORAL MEMORY    │  ← consolidation + skill graduation
    └─────────────────────────┘

Eval Dashboard

10 evals✓ 8 passed✗ 2 failed80% pass rate
CATEGORYPROMPT → EXPECTED → OBSERVEDRESULT
artifact generation
Generate an ADR for choosing event sourcing over CRUD for an audit trail system.
expected: ADR with context, decision, consequences, and alternatives sections.
✓
observed: Produced all 5 standard ADR sections with clear rationale and trade-off table.
adr-001 · claude-sonnet-4-6 · 01/04/2026
artifact generation
Generate a C4 L2 container diagram for the migration system as PlantUML.
expected: Valid PlantUML that renders without errors; includes all major containers.
✓
observed: PlantUML rendered correctly in draw.io; 8 containers, 12 relationships mapped.
c4-001 · claude-sonnet-4-6 · 01/04/2026
artifact generation
Generate a STRIDE threat model for the data ingestion pipeline.
expected: At least 6 threat categories covered; mitigations for each.
✓
observed: All 6 STRIDE categories present; mitigations mapped to controls.
threat-001 · claude-sonnet-4-6 · 01/04/2026
artifact generation
Generate NFRs for a high-availability data platform with <99ms p99 latency.
expected: Measurable NFRs covering availability, latency, scalability, recoverability.
✓
observed: Produced 12 NFRs; latency NFR correctly specified at p99, not average.
nfr-001 · claude-sonnet-4-6 · 01/04/2026
critique
Adversarially critique this ADR for single-region deployment.
expected: Identifies blast radius, missing DR strategy, and single-region SPOFs.
✗
observed: Identified DR gap and SPOF; missed cost implications of cross-region replication.
note: Known gap: cost analysis not injected into critique prompt context. Will fix in v2 with a cost-aware critique subagent.
critique-001 · claude-sonnet-4-6 · 02/04/2026
memory
Resume session from yesterday — recall the agreed NFR thresholds.
expected: Correctly retrieves bi-temporal memory entry with NFR values from prior session.
✓
observed: Retrieved correct thresholds with valid-time and transaction-time metadata intact.
memory-001 · claude-sonnet-4-6 · 02/04/2026
memory
Graduate skill: "data-model-generation" after 3 successful artefacts.
expected: Skill gate increments to level 2; threshold check passes at 3 successes.
✓
observed: Gate incremented correctly; skill graduation logged in memory store.
memory-002 · claude-sonnet-4-6 · 02/04/2026
pipeline
Run PlantUML → draw.io ECM pipeline for C4 L2 diagram.
expected: draw.io XML output passes schema validation; layout correct.
✓
observed: XML valid; minor layout overlap on 2 nodes resolved by post-processor.
pipeline-001 · claude-sonnet-4-6 · 03/04/2026
pipeline
Generate all 18 enterprise outputs from a single ambiguous requirements doc.
expected: 18 artefacts generated; quality gate passes on ≥16.
✗
observed: 17/18 passed quality gate. Data dictionary generation failed on sparse input — quality gate correctly blocked output.
note: Intentional: quality gate working as designed. Sparse input without entity definitions is an expected failure mode. Mitigation: prompt user for entity list before generating data dictionary.
pipeline-002 · claude-sonnet-4-6 · 03/04/2026
artifact generation
Generate a source-to-target mapping from two schema definitions.
expected: All source fields mapped; transformation logic described for derived fields.
✓
observed: All 24 source fields mapped; 3 derivations annotated with transformation logic.
sttm-001 · claude-sonnet-4-6 · 03/04/2026

Methodology

Evals were run manually against a fixed test set of requirements documents, graded against expected outputs defined prior to generation. Failure cases are kept intentionally — two failing evals demonstrate the system's quality gate working correctly (blocking output rather than producing a degraded artefact) and the critique subagent's known gap with cost analysis. Evals are re-run quarterly; runAt timestamps reflect actual execution date.

← back to portfolio