TT Claude Memory

project_22spike_S18_vpc_peering_anthropic_managed_trigger_2026_05_03

Properties1
descriptionThird 22-spike suite spike to close. Research-only (no kill condition). 4-trigger framework codifies WHEN to re-evaluate the 2026-05-01T09:35 A-142 lock that REJECTED option Π.ζ Anthropic Managed Agents on no-VPC-peering. context7 confirmed only unrestricted+limited networking; no VPC peering / private endpoint config; org+workspace scoping. Triggers T-1 Anthropic publishes VPC peering / T-2 compliance / T-3 cost-crossover / T-4 SLA. arch-state §13 row Q-and-A-S18 v3.31 → v3.32. T-file 371 lines. Two plan defects flagged.

4 min read

22-spike Spike 18 — VPC peering / Anthropic Managed Agents trigger CLOSED 2026-05-03T10:30 BST (research-only)

Outcome

outcome-VALIDATED (research-only spike with NO kill condition by design per plan §6 Task 18; outcome-VALIDATED is the default).

Bottom-line takeaway

The 2026-05-01T09:35 A-142 lock REJECTED option Π.ζ (Anthropic Managed Agents) on grounds of “no VPC peering / private endpoints” sovereignty-incompatibility. Spike 18 codifies the trigger conditions for re-evaluating that rejection so the next time the question surfaces, decision-quality is high without re-doing the analysis.

context7 retrieval against /websites/platform_claude_en_managed-agents 2026-05-03T10:13 BST CONFIRMED current state: only unrestricted + limited networking modes documented; no VPC peering / private endpoint config option; org+workspace scoping (no per-tenant network-isolation primitive); rate limits 60/600 req/min per organization; beta header managed-agents-2026-04-01.

The 4-trigger framework: T-1 Anthropic publishes VPC peering / T-2 compliance requirement / T-3 cost-crossover / T-4 SLA exceed self-hosted capacity. Each with measurement signal + Phase-to-act-in. T-1 is necessary-but-not-sufficient. T-2 alone forces fork (Managed Agents IF T-1 met OR doubled-down self-hosted). T-3+T-4 compose strongest commercial case. All four firing simultaneously is the unambiguous migrate signal.

If a trigger fires, migration path is ~16-24 wk wall-clock; M-5 per-firm DO state migration is hardest leg.

Key empirical findings reused at lock-time

  1. The Liveblocks REJECTION precedent (IAS arch-state Spike-E /review-plan finding) establishes the operative discipline: a managed service must offer data-residency control AND tenant network-isolation to pass IAS sovereignty review. Cloud-IAM Scaleway-EU passes (EU data-residency); Liveblocks fails (no self-hosting); by the same discipline, Anthropic Managed Agents in current beta state fails sovereignty review because it offers neither VPC peering / private endpoints nor explicit per-region data-residency configuration.
  2. Tier 2 retrieval honestly thin (top sims 0.66-0.73 vs 0.78-0.85 for established topics). Confirms Spike 18 is canonical-first-attempt territory; only v6.6 data-protection-rules.json at sim 0.724 was load-bearing prior evidence (used at §3 T-2 GDPR jurisdiction list).
  3. Phase-1 cost-row data from ε.ι suite (S2.5 owlready2 ~£2-3K + S2.9 pgvector ~£3-5K + S2.10 Cedar £15-20K) gives self-hosted ops cost shape (£20-30K Phase-1 Y1 baseline). Anthropic Managed Agents pricing NOT public in retrieved docs — §4 cost-crossover is sketch-only.

Plan defects flagged

  1. Plan §6 Task 18 Step 5 says “Stage arch-state §12 confrontation row” but Spike 18 has NO richard-task — chose §13 placement per brief recommendation (research-only architectural-decision-frame). Plan v1.9 → v1.10 patch candidate to clarify §13 vs §12 placement convention for research-only spikes.
  2. arch-state v3.28 changelog 2026-05-01T09:35 claims “5 NEW richard-tasks added” including “Anthropic Managed Agents Phase-1.5+ re-evaluation trigger” but grep of ~/davieshq/docs-personal/richard-tasks-tt.md finds no such task — Spike 18’s T-file is de-facto pre-delivery on never-landed task.

Honesty caveats

  • No live cost measurement of Anthropic Managed Agents pricing (not public per context7 surface as of 2026-05-03 BST; §4 sketch-only).
  • No partner-firm-specific compliance analysis (Anthropic Trust Center query deferred to Phase-1.5+ at trigger fire).
  • No live Managed Agents API exercise (context7 docs surface is source-of-truth proxy; if docs are stale, framework would need adjustment; mitigation: T-1 condition (a) is “blog post or release note announces VPC peering”).
  • No competitive comparison vs OpenAI Assistants v2 / Google Gemini Agents / Microsoft Foundry Agents (plan §6 Task 18 scope is Anthropic-specific).
  • Maturity is plain VALIDATED (no -WITH-METHODOLOGICAL-SUBSTITUTION nor -WITH-PROVISIONING-NOTE sub-mode); research-only spike does not face substrate / tooling / provisioning constraints.

Implications for the 22-spike suite + arch-state

  • Third spike to close (after Spike 1 + Spike 12); FIRST research-only spike with no kill condition by design.
  • arch-state §13 expanded to absorb three distinct sub-types (pre-spike infrastructure / methodological foundation / architectural-decision-frame) without further section subdivision — strengthens §13’s “suite-relatedness” semantic.
  • 4-trigger framework is starting reference for any future “self-hosted vs Managed Agents” decision; reconsideration trigger preserved at trigger fire.
  • NO change to richard-tasks.md / risk register / SKOS / CI gate / module count / primitive count.

Companion artefacts

  • T-file: ~/off-github/library/projects/inherit/T-spike-vpc-peering-anthropic-managed-trigger-2026-05-03.md v1.0 (off-github commit ae3ecc0)
  • arch-state row: inherit-v2-architecture-state.md §13 Q-and-A-S18 row (v3.31 → v3.32)
  • plan §0 Spike 18 checkbox: ticked + plan v1.8 → v1.9 + lastmod 11:00 → 11:05
  • prior-evidence aggregation: /tmp/spike-18/prior-evidence.md (4 Tier 2 queries × top-10)
  • context7 source URLs: https://platform.claude.com/docs/en/managed-agents/{quickstart,overview,environments,sessions,migration,github}
  • IK arch-state grep: ~/testatetech/docs-strategy/docs/superpowers/specs/inheritkit-architecture-state.md v1.4 changelog row 1.2 = A-142 lock with explicit “NOT Anthropic Managed Agents” rationale
  • IAS arch-state grep: ~/testatetech/docs-strategy/docs/superpowers/specs/ias-architecture-state.md §1 CRDT collab Hocuspocus self-hosted + Liveblocks REJECTED precedent
  • Memory file: this file

NEXT

  • Closure verifier bash ~/off-github/library/projects/inherit/tier-2-pgvector-library-index/22-spike-suite-closure-verify.sh --spike 18 --slug vpc-peering-anthropic-managed-trigger --day 03.
  • Report to Rich with verifier output + RED/AMBER gap counts (row 8 expected AMBER per brief — no richard-task; row 7 GREEN once active-work-log CLOSED entry added).

Reconsideration trigger preserved

If any of T-1 through T-4 fires, this T-file is the starting reference; do not assume framework is final without empirical re-baselining at trigger time.

Graph View

Backlinks