ν.β D2 — AI agent stack 6-agent loop Phase-1 vs Phase-1.5+ decomposition — CLOSED 2026-05-04
Header
Spike: D2 — ai-agent-stack-decomposition
Suite: ν.β SOTA-derisking
Outcome: outcome-VALIDATED
Target Q-NU: Q-NU-011
Target Phase Q: Q-015-ai-agent-stack-revisit (deferred to Phase 5)
Date closed: 2026-05-04
T-file: /home/richardd/off-github/library/projects/inherit/T-spike-nu-beta-D2-ai-agent-stack-decomposition-2026-05-04.md
Cascade-Q: /home/richardd/testatetech/docs-strategy/docs/superpowers/specs/2026-04-29-multi-phase-audit/current-questions/Q-NU-011-ai-agent-stack-phase-1-decomposition.md
Key finding
Kill condition NOT-MET. Phase-1 agent-loop cost estimated at £27-42K (4 agents), well below £50K kill threshold.
6-agent → L4M mapping (summary)
Post-Π.θ ★ lock (A-142), the canonical architecture is L4M 4-stage (not the old 6-named-agent loop):
| Old agent name | L4M stage | Phase scope | Cost est. |
|---|---|---|---|
| Agent 1 — Will-author | Stage 2+4 | Phase-1 | £8-12K |
| Agent 2 — Probate-author | Stage 3 | Phase-1 | £5-8K |
| Agent 3 — Asset-discovery | Stage 1 | Phase-1 | £6-10K |
| Agent 4 — Faith-tradition | Stage 1 (faith-pillar) | Phase-1.5+ | £8-15K/pillar |
| Agent 5 — Trigger-watcher | Cross-cutting async | Phase-1 partial (9/14) | £8-12K |
| Agent 6 — Acquirer-DD | Stage 4 specialisation | Phase-5 | £12-20K |
Phase-1 total: £27-42K (Agents 1+2+3+5)
Skills publication (Phase-1 cross-cutting): £10-15K additional.
Architecture stack (Π.θ ★, A-142)
- Decomposition: L4M 4-stage unified IAS+IK
- Framework: Anthropic Agent SDK programmatic (NOT LangGraph 1.0)
- Authorization: cedar-for-agents v0.4 + Microsoft AGT (April 2026; MIT; OWASP-10)
- Distribution: 9 Anthropic Skills (4 stage-skills + 5 jurisdiction-skills; faraid/HSA/halakhic Phase-1.5+)
- Sovereignty: A-113 IAS DO-per-firm preserved (Agent SDK is programmatic, not Managed Agents)
Phase-1.5+ triggers
- Faith-firm LOI received → Agent 4 activation
- HMLR Business Gateway provisioned → Agent 5 completion
- Cedar RFC #112 merged → replace SDK TemporalRule wrapper
- Anthropic Managed Agents VPC peering → re-eval IK MCP only
Key honesty caveats
- No 22-spike S14-S17 substrate (those spikes still open as of 2026-05-03) — cost estimates may shift ±20%
- Cost estimates are desktop (not sprint-validated) — analogical from comparable SDK integration work
- A-142 was registered/discarded/re-issued same day (2026-05-01) — verify current status at Phase 5 lock-time
- L4M paper benchmark on Chinese law — English-law portability reasonable but not empirically demonstrated
Q-NU-011 state
Placeholder state 1 → 3 (frontier-derisked). Awaiting Phase 5 Q-015 formal ask under refined-prompt v3.8+.
Spike 21 + Spike 22 cross-reference
- Spike 21 (μ.ε hybrid Will Update): Agent 1 £8-12K + Phase-1 UK&W scope confirmed
- Spike 22 (14-trigger taxonomy): Agent 5 £8-12K + 9/14 Phase-1 partial + 5 deferred to Phase-1.5+
Prior work anchor
Inline spike T-ai-agent-stack-spike-2026-05-01.md v1.2 is the upstream substrate (Π.θ ★ 8-option scorecard; L4M 4-stage specification; 9 Skills names; deferred_to: phase-5 per 2026-05-01T18:30 reversal).