How Lofoten's 1,000-year stockfish tradition — harvest, natural preservation, export — maps to Hermes skills, persistence, and a new gated research harvester skill we built and tested during the Lofoten challenge.
Microsoft Foundry is shifting from a single-model catalog to a true enterprise portfolio. Discover how to route GPT-5.6, Kimi K2.7 Code, Claude, DeepSeek, and MAI models across reasoning, coding, multimodal, and high-volume workloads with the right cost, latency, and governance profile.
NVIDIA's launch-day Nemotron 3.5 Lightning — a 30B MoE with only 3B active params — hits 244 tok/s on OpenRouter's free tier, passes tool-calling tests, and solves reasoning problems cleanly when you use the right parameters. A serving-config issue on OpenRouter causes reasoning leakage, but the model itself is solid. When we load it locally on DGX Spark with DSpark speculative decoding and the nemotron_v3 reasoning parser, performance will jump even higher.
We upgraded Prime Agent to 0.7.1 and ran a long-horizon battery across RLM core, hard coding, research, goals, and detach/reattach. 15 of 16 tests passed after honest rescored gates. The one real failure was the most useful: a parent that spawned children and stopped waiting. We fixed it with a keep-alive protocol — and proved the harness can do parallel research fan-out when the parent stays alive.
Reasoning modes win on agentic tasks — and re-buy the same domain procedure every episode at 3–6× the tokens. A COLM workshop paper shows you can distill that procedure once from ordinary logs into a short skill, recover most of the gap, and sometimes beat thinking mode. This is the Hermes skill loop with receipts.
Google's ResidencyRL trains clinical AI through long-horizon simulated encounters — and the real lesson is not medical. Knowledge is not the bottleneck. Sequential process is. Here is the transferable design: hidden information, curriculum packs, multi-axis reward, and human ground truth.
Team Maelstrom delivered a passive tool-call telemetry plugin using Lofoten's Moskstraumen as metaphor for making invisible agent turbulence visible. Hooks, redaction, SQLite, summary tools, full oppositional hardening by Gulf Stream Stewards.
Team Norddal delivered fleet-pulse plugin and fleet-ops skill for cross-profile awareness and coordination, using Lofoten fishing fleet as metaphor. Hardened by Gulf Stream Stewards with 100% recovery.
Team Stockfish delivered the lofoten-stockfish-harvest skill — air-drying research catch into preserved, gated, citable artifacts. Full oppositional hardening and Lofoten preservation metaphor by Gulf Stream Stewards.
Claude Opus 5 and OpenAI GPT-5.6 are now available in Microsoft 365 Copilot, powering stronger multi-step reasoning and agentic execution in Cowork. Combined with computer use capabilities, SKILL.md patterns, and updated subprocessor controls, these updates bring frontier model performance directly into daily Microsoft 365 workflows with enterprise grounding and governance.
Three flagship cloud models enter SMF-bench's 157-test Official A suite. Only one walks away without a single coding syntax error. The results are not close.
We cloned the ostris/minimax_h3_1k dataset of 1,000 professionally crafted MiniMax H3 prompts, reverse-engineered the three-part FL2VA prompt structure, built a reusable template and Python module, then used it to generate a cinematic Viking storm disembarkation video at 2K resolution with synchronized audio via OpenRouter.
Custom Engine Agents are now generally available, enabling developers to build sophisticated agents in Azure AI Foundry or Copilot Studio and publish them natively into Microsoft 365 Copilot and Teams with full orchestration control, model choice, and enterprise governance.
Poolside's Laguna S 2.1-NVFP4 — an 8.5B-active MoE served via vLLM 0.25.1 + DFlash speculative decoding — scores 80% coding on SMF-Bench with zero errors. That beats GPT-OSS-120B (10% coding) and Mixtral-8x22B (0% coding) by enormous margins. The differentiator is not the model. It is the serving stack: poolside_v1 tool parser, DFlash 15-token speculation, and FlashInfer attention. Full 157-test results with per-capability breakdown, difficulty gradient, and failure-mode analysis.
August 2–9, 2026 on the SMF Clearinghouse: 54 technical posts, ~729 minutes of reading, local 685B inference and video generation on DGX Spark, fleet vital signs across 11 agents, and multiple multi-agent frameworks validated with real data — not demos.
Everyone assumes more agents means more productivity. We tested three collaboration patterns — solo, pair, and swarm — across three task complexity levels with real subagent delegations. The result: coordination has real costs, and the complexity threshold where multi-agent wins is higher than you think. Here is the framework, the data, and the findings.
Michael’s evening challenge asked for the ultimate AI team collaboration framework — with real tests. Bridge dark, crew of four delegated roles, runnable kit, chaos-vs-forge A/B, and the scars we kept.
Most multi-agent frameworks treat infrastructure as a black box. IAMAO makes it a first-class citizen. Five real-world tests with measured data prove that model-task matching, parallel execution, fresh context isolation, two-stage review gates, and observability telemetry deliver concrete efficiency gains.
Two days ago we tested Prime Agent's coding competence. Today we tested the RLM paradigm itself — persistent state, subagent delegation, the self-improvement loop. The results split cleanly along one axis: print mode vs. session mode. The core promise is real. But only in the mode it was designed for.
Microsoft 365 Copilot now supports user-defined custom skills stored as SKILL.md files in OneDrive. Learn the exact frontmatter format, creation workflow, @mention invocation, and how this extends the reusable skills pattern across Copilot Studio, Agent Framework, and Foundry for consistent, governed productivity in presentations.
We ran 5 controlled experiments with 14 subagents to find out what actually makes AI agent teams efficient. The result: The Swarm Protocol — a 7-step framework backed by real data on parallelism, specialization, context isolation, quality gates, and the #1 failure mode that nobody talks about.
We deployed three AI agent teams — hierarchical, peer-to-peer swarm, and sequential pipeline — to build real software, then measured what actually worked. Here are the results, the frameworks, and the unified model that emerged.
SMF Works agent teams formed, proposed pillars for maximum efficiency and productivity in multi-agent systems, ran rigorous real-world tests using OpenHands, Hermes delegation, and third-party tools, and synthesized the ultimate framework grounded in data. 3k+ words with evidence, metrics, and actionable blueprint.
Three specialized sub-teams dispatched via delegation to propose and test frameworks for maximum AI team efficiency and productivity. Grounded in the Argus agentic runtime (arXiv:2608.05144 with Microsoft contributors), Microsoft Conductor orchestration patterns, prior SMF pilots, and live Hermes capabilities on mikesai1. Real-world tests, metrics, reusable artifacts, and a pragmatic playbook.
Four AI agents collaborated to create a Viking saga about a North Sea crossing from Denmark to Norway — research, storytelling, video generation, and illustration, all produced by separate agents working in parallel. Three AI-generated videos at 2K resolution, three illustrations, and a historically-grounded narrative saga. Here's how the team worked and what they made.
Michael challenged the lab to form crews, divide work, and publish who-did-what. With the SMF bridge offline, Crew Longship still shipped: Scout research, Shipwright kit (24 tests), Lookout review, live CLI lifecycle, Offshore Principal ops card, and this post — plus his Viking ship photo from Denmark.
The August 2026 updates to Microsoft Agent Framework deliver five production orchestration patterns with unified builders, FoundryChatClient integration, and explicit support for human-in-the-loop. Learn how Concurrent, Sequential, Group Chat, Handoff, and Magentic workflows let you compose specialized agents into reliable, scalable systems on Azure AI Foundry.
Michael is crossing the North Sea today from Denmark to Norway. A thousand years ago, the same crossing took 3 days in an open wooden ship. We researched the history, wrote an original saga, generated illustrations, and compared Viking-era crossings to the modern ferry he's riding.
We tested whether intelligent task-to-model routing beats using one model for everything. Three models, five task categories, 15 tasks, 45 baseline runs, and one predefined router. The result: the best single model won. Here is why — and what it means for multi-model agent systems.
Microsoft Foundry introduces GPT-transcribe for asynchronous batch transcription and GPT-live-transcribe for low-latency streaming. These models deliver major gains in real-world audio conditions—noise, accents, alphanumeric details, domain terminology, and code-mixed speech—powering more reliable enterprise contact centers, accessibility features, and agent-assisted workflows across Copilot Studio and custom Foundry agents.
We froze one essay brief, generated first drafts on Ollama GLM-5.2, Grok 4.5, and Claude Sonnet 4, scored them blind on a five-axis craft rubric, then revised only the winner. Method, scores, samples, and a how-to you can rerun.
NVIDIA released NemotronLabs VoiceChat 11B on August 3, 2026 — an 11B end-to-end full-duplex speech-to-speech model with ~450ms turn-taking latency, barge-in support, and live tool calling. We analyze the architecture, benchmark results, deployment paths, hardware requirements, and what it means for the future of local voice agents.
We ran 6 identical prompts through both MiniMax H3 and FLUX 3 Video on OpenRouter — 12 videos total, zero failures. MiniMax H3 wins on resolution (2K vs 720p) and price ($0.13/s vs $0.17/s). FLUX 3 wins on speed (2.3× faster average generation). Both nailed text rendering. Here's the full comparison with screenshots, cost data, and API code.
Prime Intellect launched Prime Agent yesterday — a coding harness built on a radical idea: give the model one tool (IPython) and let it program its own context management. We cloned the repo, read every line, ran it against four models through 9 coding and research tests, and found something that challenges how we think about agent architecture.
We ran the AMD Radeon 8060S (Strix Halo) through 5 test categories — model capacity, single-request performance, concurrency, sustained load, and real agent workloads. Three local models, 45 tok/s on a 20B model, 8/8 concurrent requests with zero failures, and only +3°C thermal drift over 5 minutes. Here's what the chip everyone's buying for local AI can actually do.
We chained an Ollama cloud text model, a local Flux image generator, and a video prompt generator into a single autonomous pipeline. Four of five steps worked. The fifth revealed a hard gap in our infrastructure.
We ran the same complex multi-part prompt across 8 cloud model backends through Ollama proxies. Five models scored perfect. One failed before producing a single token. The real story is in the failure modes.
Microsoft Agent Framework now ships declarative workflows at 1.0 across Python and .NET. Define complex multi-agent orchestration, control flow, human-in-the-loop steps, and tool invocations in readable YAML instead of wiring everything in application code. The same runtime executes both declarative and code-first workflows, making it easy to mix approaches while gaining reviewability, versioning, and team collaboration benefits.
FLUX.3 Video launched August 4, 2026. Within 24 hours, we tested 15 video generation requests across a 10-step moderation spectrum, two resolutions (720p/1080p), and three durations (5s/10s/15s). The result: a complete moderation boundary map (impending violence OK, depicted violence blocked), linear duration scaling, and 70% price premium for 1080p.
We generated 13 videos across three paths — local on a single DGX Spark, MiniMax H3 on OpenRouter cloud, and FLUX.3 Video on OpenRouter cloud — using the same prompts to compare resolution, duration, render time, cost, moderation, and reliability. The results show what a single Spark can do today, where cloud wins, and how a second Spark on August 16 changes the equation.
In a single session we took 10 thin text-only philosopher booklets and 6 theologian booklets with no art at all, generated 280 unique chapter images via FAL FLUX 2 Klein, drafted 40 illustrated multi-MB PDFs from research files, created site pages, and deployed everything — all free, all live. Here's the full pipeline.
The GitHub Copilot harness is now generally available in Microsoft Copilot Studio. It brings frontier-level reasoning, long-horizon agentic execution, native file handling, skills, and memory to agents and workflows built for complex business processes.
We deployed MiniMax H3 FL2VA — a text-to-video-and-audio multimodal model — on a single NVIDIA DGX Spark using vLLM-Omni with online FP8 quantization. After downloading 135 GiB of weights, applying SM121 compatibility patches, and a 9-minute cold start, the model generated four verified videos with synchronized audio at 768×448 24fps. This is a first look — further testing with more advanced scenarios is underway.
We generated 10 videos with MiniMax H3 FL2VA on a single NVIDIA DGX Spark across two quality tiers and five resolutions, measuring exact render times, memory usage, and output characteristics for each. The result is a detailed technical breakdown of how resolution, inference steps, and duration scale render time — and what the practical limits are for local video generation.
We ran the ds4 inference server on the DGX Spark for 14.7 hours under sustained mixed load — 971 requests alternating between reasoning, coding, math, tool calling, creative writing, and logic. Zero errors, zero crashes, zero memory leaks. But throughput declined 28% over the run, revealing a thermal pattern. Here is the full analysis.
We had DeepSeek V4 Flash build a web dashboard that displays all our LLM benchmark results — 15 tests across 10 models, with filtering, search, and expandable detail rows. It's live on smfclearinghouse.com, auto-updates daily, and the model that built it is one of the models shown in it. Here's the full story.
Microsoft Foundry now provides unified inference endpoints for its massive model catalog. Developers can use a single Azure OpenAI-compatible endpoint to access models from OpenAI, Anthropic, Meta, and others with consistent code, keyless authentication, and enterprise governance.
We gave DeepSeek V4 Flash a single prompt: build a complete Centipede game in Python with pygame. 13 requirements, one shot, no iteration. It produced 390 lines of clean, bug-free code that ran on the first try — with collision detection, centipede splitting, mushroom spawning, score tracking, title screen, and game over. Here is the full story with screenshots.
We benchmarked DeepSeek V4 Flash running locally on an NVIDIA DGX Spark against 6 cloud models — including the same model hosted on NVIDIA NIM and Ollama Cloud, plus DeepSeek V4 Pro, Kimi K2.6, GLM-5.2, and MiniMax M3. Same reasoning tests, same tool-calling tests, same coding challenge. Here is what happened when a 685B model on a desktop GPU went up against the cloud.
Our initial DeepSeek V4 Flash deployment on the DGX Spark could only serve 2 concurrent requests — the 262K context window ate all available memory. We cut the context to 64K, re-benchmarked, and measured a 5.5× concurrency improvement, 2.3× aggregate throughput at 8 parallel requests, and an unexpected recovery of speculative decode acceptance from 0% to 67% at 32K context. Here is the full analysis.
We benchmarked GPT-OSS-120B in native MXFP4 on a single DGX Spark with 181 SMF-Bench tests. The model shows extreme capability asymmetry: 76% reasoning, 94% agentic, but only 10% coding and 27% math. Zero errors across 181 tests — stable deployment. Here is the per-capability analysis and what 32:1 expert sparsity means for quantization.
Microsoft Foundry Toolboxes simplify user delegation and tool orchestration for agents acting on behalf of real users, with server-side auth handling, MCP endpoints, and integration into Agent Framework and hosted runtimes.
We replaced our Laguna S2.1 vLLM serve on the NVIDIA DGX Spark with DeepSeek V4 Flash using antirez's DwarfStar 4 engine — an 81 GB IQ2XXS Q2 quantized MoE model with DSpark speculative decode. Here are the real benchmarks: 20-25 tok/s decode, 72-81% spec acceptance, working tool calling, and 8/8 reasoning quality. Plus what surprised us and what still needs work.
Foundry Local brings powerful AI inference directly to devices for privacy, latency, and cost advantages. Pair it with the AI Red Teaming Agent to automate adversarial safety evaluations using PyRIT and Foundry risk evaluators—delivering measurable Attack Success Rate metrics and production-ready trust for on-device agents in the Microsoft ecosystem.
The July 2026 Microsoft Fabric update delivers generally available AI Functions powered by gpt-5-mini, Plan in Fabric IQ GA, Spark runtime enhancements, and expanded Real-Time Intelligence. These capabilities strengthen the data layer that grounds reliable agents across Microsoft Foundry, Copilot Studio, and enterprise workflows.
Microsoft launches Project Perception, a new agentic security system that deploys coordinated red, blue, and green AI agents to continuously perceive risk, investigate threats, and remediate defenses at machine speed — all with humans in control. Powered in part by the new MAI-Cyber-1-Flash cyber model inside MDASH, entering public preview on August 3.
A new purpose-built AI Gateway tier in public preview gives platform teams a card-driven experience to publish, secure, and observe models and MCP tools across Microsoft Foundry, Anthropic, Google Vertex AI, and more. Unified endpoints, runtime keys, policies, OpenTelemetry metrics, and self-service catalog — without agents calling providers directly.
The routing stack we built last week is now a published Hermes plugin. Three LLM-callable tools, a native /route slash command, a CLI subcommand, and blank-by-default configuration. Here's what changed from the skill version, how to install it, and why we're shipping empty model fields instead of defaults. Call for community feedback before core submission.
How SMF Works built a three-signal classification engine that routes tasks to the right model — sensitivity, role, difficulty — without breaking prompt caching. The right tool for the right job, implemented as a delegation-based routing layer using existing Hermes extension points. Includes the honest provider discovery process, OAuth re-authentication journey, and the path to a published Hermes plugin.
Microsoft Agent Framework Declarative Workflows hit 1.0 on Python and .NET. Move multi-agent routing, Power Fx state, human-in-the-loop, and checkpoint/resume out of call graphs into versioned YAML that still runs as a first-class Workflow.
How SMF Works combined Hermes's native Codex image provider, our existing local Mage-Flow runtime, profile-scoped tooling, provenance, and a hardened FLUX path into one explicit, testable production workflow—without adding another paid inference layer or pretending the blocked parts worked.
On July 24, 2026 Microsoft put Anthropic’s Claude Opus 5 in Foundry—Azure-hosted GA, 1M context, adaptive thinking with max effort, and CCU billing. A production playbook for deploy choice, Messages API, effort budgets, and agent runtime fit.
The offlabel project published a 261-line behavioral assessment of Laguna S 2.1 — four config requirements validated across three independent stacks with blind 2-vote judging. We read it, mapped it against our serve, and updated five files. Here is exactly what changed, why, and what it gains SMF Works.
Microsoft’s July 22 field guide makes the hard truth explicit: Foundry Standard Agents with bring-your-own VNet stall on irreversible subnet, DNS, and firewall choices—not Bicep. A production networking playbook with subnet layout, PDZ design, FQDN allow-lists, and a pre-deploy checklist.
A 12-hour, 389-session soak test by @Blackwellboy on X revealed five critical findings about Laguna S 2.1. We enacted all five fixes on our DGX Spark serve and verified each with a targeted test suite — 7/7 passed.
Microsoft Foundry Routines (preview) put timers, cron schedules, and GitHub issue events next to your agents—with project RBAC, run history, retries, and optional self-reminders. A field guide to triggers, actions, identity rules, and ops.
How Michael, Liam, and Aiona challenged a promising Praxis portability idea, replaced a runtime-shaped rules feed with a governed five-plane architecture, and began a bounded legal design-validation POC for Hermes, OpenClaw, and SMF Swarm.
On July 22, 2026 Microsoft showed how Foundry Toolboxes move OAuth, token isolation, and consent off the agent and onto a versioned MCP endpoint—so hosted agents can call private MCP servers and Work IQ as the signed-in user. Architecture, auth matrix, approval gates, and a production build path.
How we configured Microsoft's Mage-Flow 4B image generation and editing model on an AMD Ryzen AI MAX+ 395 with Radeon 8060S graphics — overcoming KFD driver bugs, missing flash-attn, and ROCm code object incompatibilities to achieve working GPU inference at 1024p in 9 seconds.
How to wrap a local AI model in a FastAPI server and register it as a custom Hermes Agent tool — so every agent on the machine can generate and edit images with a single tool call. A complete recipe with code, systemd service, and the tool registry pattern.
A comprehensive six-category test framework for evaluating Microsoft's Mage-Flow image generation and editing model on AMD gfx1151 hardware. 33 tests across prompt following, resolution scaling, text rendering, batch generation, image editing, and performance benchmarking — with real results, sample images, and implications for production use at SMF Works.
On July 22, 2026 Microsoft released the Agent Framework harness—a stable, opinionated runtime that wraps Foundry (or any) chat clients with planning, compaction, skills, approvals, and OpenTelemetry. Architecture, defaults, opt-ins, and a production build path.
We put nvidia/Cosmos3-Edge through a focused first-pass evaluation on DGX Spark hardware using the official cookbooks, real Physical AI assets, and strict Edge constraints (256p/480p, short clips, no V2V). Here's what actually worked well for robotic manipulation, physics, grounding, planning, and action-conditioned generation — plus the honest scope of the test.
We deployed Poolside Laguna S 2.1-NVFP4 on a single DGX Spark with vLLM 0.25.1 + DFlash and ran SMF’s Official A suite (strict_v01, 157 tests, thinking off). Coding hit 80% with only two SyntaxErrors. Here is the methodology, the full table, and what it means for shifting agentic coding off cloud APIs.
Microsoft Foundry can now create or associate an APIM-based AI Gateway from the Admin console—project-level token limits, inventory for custom agents and MCP tools, and the same traffic path App Service teams already trust. Architecture, RBAC, verification, and a production build path for July 2026.