What's New in the Engine Fleet
Benchmarked against: Anthropic — What's new in Claude Last updated: 2026-03-05
This page tracks significant engine fleet changes. For full release history, see Release Notes.
March 2026
Cloud UB unified as Single Source of Truth (2026-03-01)
Cloud UB (Cloudflare D1 + Vectorize) declared as the only Universal Brain. All ships (SS1, SS2, SS3) unified on Cloud UB. Local UB entries (2,280) archived with high-quality items to be migrated.
Impact on engines: All engines now read from and write to Cloud UB exclusively. Local UBI tools still route through Cloud UB for data operations.
Dispatch Worker migrated to Cloud UB (2026-03-02)
The SS1 Dispatch Worker (which executes Work Orders via low-cost engines) updated to use Cloud UB instead of local UB. This enables any ship to dispatch and monitor work orders.
Docs site launched (2026-03-05)
SuperPortia Agentic AI Docs site goes live internally. Engine documentation, selection guides, and pricing information now available as structured reference pages.
February 2026
Dual Agentic Lines strategy confirmed (2026-02-27)
Captain confirmed two parallel agentic lines — NOT pick one:
- Line 1: Claude Agentic (Agent SDK + Claude Code) ~30%
- Line 2: LangGraph self-built ~70%
Both lines are truly agentic (autonomous perceive-reason-act-learn), with mutual disaster recovery.
Engine selection principle clarified (2026-02-27)
Captain decision: "CP value is NOT cheapest possible but minimum cost that gets the job done RIGHT."
| Task level | Engine |
|---|---|
| Trivial | Groq (free) |
| Standard | Gemini / DeepSeek (cents) |
| Important | Gemini with citations |
| Critical | Claude |
| Architecture | Opus |
Groq tool-calling limitations discovered (2026-02-21)
Groq Llama 3.3 70B uses XML format for tool calling instead of JSON. This breaks LangGraph nested agent-as-tool mode.
Mitigation: supervisor.py implements automatic flat/nested mode switching:
- Groq, DeepSeek, Mistral, Zhipu → flat mode (direct tools)
- Gemini, Claude → nested mode (agent-as-tool workers)
LangGraph ecosystem installed on SS1 (2026-02-26)
Verified installation in .venv: langgraph 1.0.8, langgraph-checkpoint-sqlite 3.0.3, langchain-mcp-adapters 0.2.1.
Token baseline measured: single Q&A with 1 tool call ≈ 20,677 input + 102 output tokens (Sonnet). Tool descriptions alone = ~4,103 tokens for 35 tools.
Groq hallucination incident (2026-02-22)
Ino Scout Daily using Groq/DeepSeek produced fabricated S-grade intel claiming "LangGraph 2.0 Released." Actual PyPI latest was 1.0.9.
Policy change: All intel from low-cost engines must be verified against package registries (pip/PyPI/npm) before UB ingestion, especially for perishable knowledge.
Engine fleet status
| Engine | Status | Notes |
|---|---|---|
| Claude (Opus 4.6) | Active | Primary engine, Max Plan |
| Claude (Sonnet 4.6) | Active | Shared quota with Opus |
| Claude (Haiku 4.5) | Active | Fast/cheap Claude option |
| Groq (Llama 3.3 70B) | Active | Free tier, flat mode only |
| Gemini (2.5 Flash) | Active | Search + embeddings |
| DeepSeek (R1/V3) | Active | Reasoning tasks |
| Mistral | Active | European alternative |
| Zhipu (GLM-5) | Active | Chinese NLP specialist |
Related pages
| Page | Relationship |
|---|---|
| Engine Overview | Full engine catalog |
| Release Notes | All release notes |
| Engine Migration | Switching engines |