Research Pass 2 — Open Questions Closed

As of: 2026-06-04. Closes the seven items left open after pass 1 (A1, A2, A3, D1, D2, D4, D6). Each finding is anchored to primary or near-primary sources; treat the conclusions as directional and re-verify before locking architectural decisions.

Companion to PENDING.md and RESEARCH-PICKS.md.

A1 — Closed-tool architecture clues (Decagon / Sierra / Moveworks / Cresta / Aisera)

Decagon — Agent Operating Procedures (AOPs): structured procedures in plain English that compile to API calls + knowledge-base lookups + escalation criteria. A built-in Routing module decides AI-vs-human per inquiry. Compliance posture: SOC 2, ISO 27001, GDPR, HIPAA. Watchtower does real-time QA + guardrails. Strengths: knowledge-base-heavy SaaS support with an audit-trail-friendly behavior contract.

Sierra — Constellation architecture routes across multiple LLM providers with model-level guardrails to reduce hallucination. Voice-first deployment is the headline. Ghostwriter (Mar 2026) builds production agents from SOPs, call transcripts, whiteboard sketches, or plain-English descriptions.

Sources:

Decision implication for TENET:

Both lean on multi-model routing under one orchestrator + structured procedure DSL. TENET’s @tenet/router (C3) confirms multi-model routing as table stakes. The procedure-DSL idea is interesting but not blocking — we get most of it from prompt-templated state-machine nodes plus the constitutional-principle registry.
Audit-trail-friendly behavior is a first-class need — our OTel agent.* attributes already cover this.
Routing-as-explicit-decision: our adaptive router (cheap-vs-flagship) is one form; an AI-vs-human escalation path is a separate concern we should make explicit in C10 (community-bot) HITL hooks.

A2 — Vector-store cost-perf at 100M+ vectors

Synthesized from a 12-DB benchmark on 100M × 768-dim vectors (identical hardware):

DB	100M-vector cost / mo	Notable
Turbopuffer	$800	Object-storage-native; cold-query 300-800ms; lacks hybrid search
Qdrant Cloud	$5,000	Lowest latency — p50 4ms, p99 25ms among purpose-built
Weaviate	$3,500	Hybrid search built in
Pinecone	$7,000	Mature; high cost at scale
pgvector	n/a	Buckles above 50M on a single node — index builds > 2 hours, p95 > 200ms without careful tuning

Sources:

Decision implication for TENET:

pgvector remains the default for phase 1 (single-tenant, <50M vectors). It is NOT appropriate for enterprise multi-tenant > 50M.
Phase 2 default: Qdrant (purpose-built, lowest latency, mature hybrid search).
Cost-sensitive multi-tenant deployments: Turbopuffer adapter as a phase-3 option for archive-tier corpora where 300-800ms cold-query latency is acceptable.
The @tenet/retrieval VectorStore interface is already designed to swap; we ship pgvector + Qdrant adapters and document the >50M cliff in the per-package READMEs.

A3 — MCP enterprise adoption signal

Production adoption is real and material:

Anthropic (Mar 2026): 10,000+ active public MCP servers; 97M monthly SDK downloads (Python + TypeScript).
Registry growth: 7.8× year-over-year; 1,200 servers (Q1 2025) → 9,400+ (April 2026); month-over-month growth +18%.
Production deployments at named Fortune-500s: EY, JPMorgan, Salesforce, Reddit, Amazon, Block, Bloomberg orchestrate MCP in production.
Workato Enterprise MCP: “trusted by half of the Fortune 500.”
Stacklok 2026 software report: 41% of surveyed software orgs in limited or broad production with MCP servers.

Sources:

Decision implication for TENET:

“MCP-native” stays a launch-grade differentiator, not a phase-2 nice-to-have. The market is large enough that operators expect it.
Top enterprise blocker is security: audit trails, SSO-integrated auth, gateway behavior, configuration portability. Our @tenet/core brand-typed Secret<T> + PathHandle + filter-key validation address two of those; we still owe an MCP gateway pattern (phase 3).
Worth adding a connectors/mcp/ package to the roadmap that wraps the MCP TypeScript SDK behind our ToolUse interface.

D1 — Claude Agent SDK + Opus 4.8 + Haiku 4.5 current surface

Opus 4.8 (released 2026-05-28) — confirmed picks in BENCHMARKS.md remain accurate. Key SDK-relevant updates:

1M-token context by default on Claude API, Bedrock, Vertex AI; 128k max output tokens.
Prompt cache minimum lowered from previous tier to 1,024 tokens — many prompts that couldn’t cache on 4.7 now cache for free. $0.50/M input tokens for cache reads (90% discount).
Mid-conversation system messages — system entries accepted inside the messages array; lets you update instructions mid-task without breaking the prompt cache or routing through a user turn.
Dynamic workflows — Opus 4.8 plans a task and runs hundreds of parallel subagents in a single session. Anthropic positions this as the path to codebase-scale migrations.
Effort control — defaults to high; same token spend as 4.7’s default with better quality.
Adaptive thinking — only reasons when needed; cuts wasted thinking tokens on bimodal workloads.

Sources:

Decision implication for TENET:

Lower cache threshold changes our cost math — many prompts that were too short to cache now cache. Update BENCHMARKS.md cost-per-resolution model accordingly when we land the cost meter integration.
Mid-conversation system messages + adaptive thinking suggest our @tenet/policy composer should emit a delta system message on intent change instead of rebuilding the whole prompt. Track as a phase-2 polish.
Dynamic workflows / 100s of subagents map cleanly to our @tenet/router speculative-agent pattern (C3). We can lean into Anthropic’s primitives for the speculative path.

D2 — MCP spec changelog since 2026-06-03

Release candidate (2026-06-04 status): the 2026-07-28 spec release is the largest revision since launch. The RC is live now.

Key changes:

Stateless core — scales on ordinary HTTP infrastructure (no special transport).
Streamable HTTP requires Mcp-Method + Mcp-Name headers (SEP-2243) — load balancers / gateways / rate-limiters can now route on the operation without inspecting the body.
ttlMs + cacheScope (SEP-2549) on list and resource read results — HTTP-Cache-Control-modeled freshness signals. Clients can share tools/list responses across users when safe.
OAuth alignment: clients must validate the iss parameter on authorization responses (RFC 9207, SEP-2468) — mitigates mix-up attacks in MCP’s many-server pattern.
OpenID Connect application_type declared during Dynamic Client Registration (SEP-837) — fixes the desktop/CLI client → “web” defaulting bug.
Extensions: server-rendered UIs (MCP Apps), long-running work (Tasks).

Sources:

Decision implication for TENET:

Pin our MCP integration to a specific spec version (probably 2026-07-28 when it ships) and document it in CHANGELOG. Pre-July deployments should target the RC.
Take a hard dependency on Mcp-Method + Mcp-Name header routing in our rate-limit scheduler — these turn MCP tool calls into platform-routable traffic at the @tenet/rate-limit layer.
The iss validation and application_type rules apply to any MCP-OAuth flow we ship — bake them into the future connectors/mcp package.

D4 — OSS framework refresh (Letta / Mastra / Pydantic-AI / Inkeep / OpenAI Agents SDK)

Framework	2026 positioning	Notable
Mastra	De facto TypeScript choice	19k+ stars; 300k+ weekly downloads; workflows + HITL primitives + Next.js integration
Pydantic AI	Best Python-first for structured outputs / validation / DI	Clean app architecture
OpenAI Agents SDK (TS)	Next-best TS choice after Mastra
Letta	Niche — persistent-memory assistants
Inkeep SDK	TS extension to no-code platform; MCP + A2A protocols; deep customization
LangGraph	Still strongest mindshare; complex flows	We design out the ecosystem’s CVE-class mistake categories by construction (see SECURITY.md); no advisory attributed here to any named project
CrewAI	Multi-agent role-based; simpler than LangGraph
Google ADK	Multi-language flagship (Python + TS + Java + Go)

Sources:

Decision implication for TENET:

Mastra is the direct competitor for the TypeScript agent niche. Our differentiators (verification-by-default, source-grounded URL allow-list, hardened-by-construction security stance, all-surfaces-one-config, OTel GenAI native) remain meaningful — Mastra leans on workflow + HITL polish, not on hallucination defense.
The Inkeep precedent (TS SDK + MCP + A2A) suggests we should hint at A2A protocol support in the roadmap if we want enterprise integration parity.
We should explicitly compete with Mastra in the README (it’s not in our current comparison table). Add it in a later commit.

D6 — Reranker head-to-head (Cohere / Voyage / BGE / Jina)

Public leaderboard snapshot (ELO + nDCG@10):

Reranker	Top metric	License
Zerank 2	1638 ELO (top)	hosted
Cohere Rerank v4.0 Pro	1629 ELO	hosted
Voyage Rerank 2.5	Fastest (~595ms p50) + quality	hosted
Cohere Rerank 3.5	~603ms p50; less favored by LLM judge	hosted
Jina Reranker v3	61.94 nDCG@10 BEIR; 131k ctx; listwise (64 docs/pass)	Apache-2.0
BGE-Reranker-v2-M3	Lightweight, multilingual, fast	Apache-2.0
Qwen3-Reranker-4B	69.76 MTEB-R / 75.94 CMTEB-R	Apache-2.0

Sources:

Decision implication for TENET:

Revised default ordering (replaces tentative pick in RESEARCH-PICKS.md):

Voyage Rerank 2.5 — primary hosted: fastest + competitive quality.
Cohere Rerank v4.0 Pro — secondary hosted: 1629 ELO, broader provider footprint.
Qwen3-Reranker-4B — primary self-host: strongest open-weights as of 2026-06-04, Apache-2.0.
Jina Reranker v3 — secondary self-host: listwise + huge context (specialty long-doc workloads).
BGE-Reranker-v2-M3 — practical baseline: small, multilingual, easy ops.

Our Reranker interface in @tenet/retrieval is unchanged — each is a 4-line adapter. Update RESEARCH-PICKS.md reranker section when this ships.

Net deltas vs prior research-picks

Reranker primary changed: Cohere 3.5 → Voyage Rerank 2.5 (faster + better quality at MTEB).
Self-host reranker primary changed: BGE-Reranker-v2-M3 → Qwen3-Reranker-4B (stronger benchmarks, same Apache-2.0).
Vector-store phase-2 default: Qdrant confirmed (lowest latency at 100M).
Vector-store budget-tier: Turbopuffer added as a phase-3 archive-tier adapter target.
MCP: launch-grade differentiator confirmed, not deferred. Pin 2026-07-28 spec.
Direct competitor: Mastra. Worth explicit README mention.
Opus 4.8 cache threshold: 1024 tokens — update cost model.