iCreatAI News
Back to main site
Skip to content
#Deployment
Nov 11Wed
  1. TechCrunch · AI38

    Plurai wants to help AI agents work better using the same approach that helped get self-driving cars to roads

    Plurai, founded by former Nexar engineers Ilan Kadar and Elad Levi, has raised $13 million in seed funding led by Team8 to build a simulation layer for AI agents, using synthetic data based on a company's own systems to prepare models for edge cases. The platform creates a digital twin of a company's knowledge base and tools, maps possible agent flows, and uses that to evaluate and fix failure modes, working alongside tools like Claude Code, Codex, Copilot and Cursor. Kadar says the company's real promise is post-training small open-weight models tailored to customers, with tests alongside Nvidia and Nebius showing performance above mid-tier frontier models at lower cost.

  2. Hacker News AI58

    Show HN: Tensward – Local AI inference optimisation

    Tensward is an Apache-2.0 tool that measures how your LLM is served on your own machine by starting its own copy of vLLM or llama.cpp on a local port, sending a realistic workload, and reading engine and GPU metrics. It offers one flag change at a time, measures it, compares answers before and after, and keeps the change only if it is faster and the answers pass your check; it never stops or reconfigures your running server. It supports NVIDIA and Apple Silicon hardware, and the post includes two L4 case studies showing gains such as +40% requests/s and first word −47% on llama.cpp, and first word 19 s → 0.30 s on vLLM.

  3. Hacker News AI62

    Show HN: Hubzoid – Turns a folder of Markdown into an AI agent your team can use

    Hubzoid is an Apache-2.0 tool that turns a folder of Markdown files—knowledge, rules, skills and agent instructions—into a shareable AI agent usable in chat, repeatable workflows, or via MCP in Claude Code, ChatGPT or Codex. It runs on an existing Claude Code or Codex login (or a model API key), supports team sign-in, per-person and per-agent access, approvals before consequential actions, and a record of every run. The starter hub is a fictional shop whose refund rule in knowledge/house-rules.md can be edited to see answers change, with evals to check they stay correct.

  4. Mastra Blog58

    Introducing MCP Tasks in Mastra

    Mastra now supports MCP Tasks in its MCPServer and MCPClient, letting long-running tool calls run in the background so the agent loop stays unblocked while the client polls the server for status (working, input_required, completed, failed, or cancelled). Tools can be configured to run as tasks with "always", "auto" (after syncWindowMs, 2 seconds by default), or "off", and tasks can still use suspend() for human input. The feature requires @mastra/[email protected] and @mastra/[email protected] or later and is still experimental, so the API may change between releases.

  5. Hacker News AI38

    Modernizing NetHack's C engine with .NET MAUI, Avalonia and a 31-tool AI advisor

    Sound Mind Games (Hyvän mielen pelit ry), a Finnish nonprofit, describes how GnollHack wraps NetHack's roughly 500,000-line C99 engine as a native library behind a shared C# client, migrated from Xamarin.Forms and SkiaSharp to .NET MAUI by April 2025 and extended to Linux, the Steam Deck and macOS in 2026 via Avalonia.Controls.Maui. The game's opt-in Gnoll Overseer advisor embeds a web app that reads live game snapshots from the C engine and uses 31 tools (20 server-side, 11 on-device) to look up facts, and GnollBench grades models with the production prompt and tools using a blind two-provider judge panel, a geometric-mean score, a 25-point critical-error cap and a 50-point honesty floor. The post also covers the October 2026 nine-model benchmark results and the team's agentic workflow, in which coding agents write most C# and TypeScript code while the C engine, art, music and voiceovers remain human-made.

  6. Hacker News AI58

    Show HN: I checked 10 AI-written patches to real C++ projects – 6 didn't hold up

    Unfix is an independent verifier for C/C++ and firmware bug fixes that reproduces the bug on the base version, applies the fix with sanitizers on, checks other tests still pass, then reverts the fix to confirm the bug returns, returning VERIFIED, REJECTED or NOT VERIFIABLE without an LLM in the verdict. In a public benchmark of 10 AI-generated patches (from Codex CLI) against real bugs in fmt, spdlog, zlib-ng and cpp-peglib, judged against maintainers' hidden reference tests, Unfix verified 4 and rejected 6. It runs as an MCP tool in agents like Claude Code and Codex, as a GitHub PR check (demo repository only), in the terminal, and on STM32 NUCLEO boards via ST-Link.

  7. Hacker News AI36

    Show HN: Episteme AI – Self-healing roadmaps with paywall and dead-link checks

    Episteme AI is a Next.js app that turns career goals into time-bounded, prerequisite-ordered weekly roadmaps built from free public courseware, using Gemini 2.0 Flash with a deterministic 52-week fallback. It uses Cheerio to detect paywalls and dead links via Schema.org JSON-LD and HTTP checks, then swaps in verified free alternatives, with approval-gated calendar and audit-link dispatch backed by SQLite idempotency keys. A live demo is available at episteme-ai-one.vercel.app.

  8. Hacker News AI58

    Show HN: Outlier, AI without the data center, private and offline on your Mac

    Outlier is a notarized macOS app that runs local AI models entirely on Apple Silicon, keeping chats, files and code on the user's disk unless web search or a cloud model with the user's own API key is enabled. It offers six model tiers from a free 4B Nano up to a 397B Plus tier, with a one-time $49 Pro license covering all tiers and features, and requires M1 or newer on macOS 15+. The post includes benchmark tables, socket-sampling evidence of local-only inference, and an explicit list of limitations, including that it does not yet match top cloud coding agents.

  9. Hacker News AI40

    Model Desk – A local-first multi-provider AI workspaceShared via URL-sharer

    Model Desk is an open-source, MIT-licensed Linux desktop app built with Rust and Tauri that connects to Lumo via lumo-tamer, DeepSeek, and any OpenAI-compatible endpoint, with an OpenCode coding agent, local history, attachments with local OCR, tools with permissions, memory, and a Bonsai image studio. Version 0.4.1 adds automatic Bonsai backend start, Lumo Tamer, DeepSeek, and the OpenCode agent, and is distributed as an AppImage or .deb for Linux x86_64. The app stores data in ~/.config/model-desk and reports no telemetry.

  10. Hacker News AI42

    Show HN: Halyard – Mac and iPhone app for AI agents in tmux, nothing to install

    Halyard is a Mac, iPhone and iPad app that opens tmux sessions on your own servers over SSH, so AI coding agents like Claude Code, Codex and Gemini keep running when you close the lid or lose connectivity. The Mac app renders with libghostty on the GPU, supports full tmux history search, SFTP file browsing and Keychain-stored keys, and connects directly with no relay or account. It requires a reachable SSH server with tmux, macOS 15 or later, and iOS/iPadOS 17 or later, with a 14-day trial and Halyard Pro at $24.99.

  11. vLLM Blog58

    Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD

    vLLM details its Encoder/Prefill/Decode (EPD) disaggregation work, including metadata-only requests that let downstream instances reuse remote embeddings without repeating media preprocessing, a dedicated encoder-only Model Runner V2 path, and remote embedding transfer via NIXL and Mooncake. It also covers cross-encoder output reuse through Mooncake Store and orchestration support in vLLM Router, SMG, and llm-d, with core capabilities in vLLM v0.32.0 and DeepSeek V4.1 Flash metadata-only support on main via #62802. A DeepSeek V4.1 Flash benchmark on GB200 found EPD delivered 6.6–9.5% higher per-GPU throughput and 44–47% lower P90 TTFT than PD at concurrency 64–128, but lower per-GPU throughput at concurrency 16.

  12. Hacker News AI34

    We Run a Bootstrapped AI Startup on Almost Zero Operating Costs

    A former co-founder and CTO of a bootstrapped AI ed-tech startup describes how the team covered most product infrastructure through startup programs, credits and free tiers rather than paying for each SaaS service, including Sarvam LLM credits, Deepgram and ElevenLabs speech credits, Cloudflare, PostHog, Sentry, Neon and others. He details the mock-interview STT/TTS loop, concurrency limits across providers, and fallback routing, estimating roughly 235 concurrent interviews on paper with 20% headroom across three STT providers. He concludes that credits are not an architecture, since they expire and limits change, so integrations were kept swappable with fallbacks for critical parts.

  13. Hacker News AI18

    I'd rather use AI than Ansible

    The author suggests that for provisioning a small number of bare metal servers, using prompts and natural language might be simpler than writing YAML, and asks for a reality check. The post links to a Hacker News discussion thread.

  14. Hacker News AI58

    Apple's Claris Testing 'Harbor' App for Local AI on Macs

    Apple's Claris subsidiary is testing Harbor, an early-access app that turns Apple silicon Macs into a private AI inference server for open-weight models, with data staying on local hardware. Harbor can cluster multiple Macs over Thunderbolt 5 to pool unified memory for larger models and higher concurrency, using RDMA support added in macOS Tahoe 26.2, and requires no manual configuration or command line. The App Store listing has been removed, and the sign-up page asks for a company name, suggesting a business focus.

  15. Hacker News AI22

    Brazil's AI Development Strategy Needs an Edge Layer – TechPolicy.Press

    TechPolicy.Press argues that Brazil's AI debate is overly focused on frontier models and hyperscale data centers, while edge AI—local inference on small, specialized models—better fits the country's industrial and employment needs. The article cites Brazil's 2.3 billion reais ($444.2 million) investment in supercomputing projects with Huawei and iFlytek, and an FGV study estimating that only about 4% of jobs at a 100-megawatt data center become permanent operations and maintenance roles. It points to China's sector-specific industrial AI approach and the CESAR/UCB Power battery-inspection system in Manaus as models, and criticizes the Redata tax regime for excluding Simples Nacional companies.

  16. MarkTechPost62

    Fish Audio Releases Live 3, a Full-Duplex Voice Model That Reacts in 186 ms

    Fish Audio has released Live 3, a full-duplex voice model for real-time voice agents that keeps listening while it speaks and learns turn-taking from real human conversations rather than hand-tuned pipeline rules. Fish Audio says it can decide to respond in as little as 186 ms for short reactions like "mm-hmm," though no benchmark scores or latency distribution have been published. It is available on the Fish Agent Platform, via the Fish Audio API, and on-premises for enterprises, priced at $0.035 per minute until December 31, 2026 (then $0.07), and supports English only.

  17. Hacker News AI42

    The Unexpected AI Stack: C# + .NET

    This post argues that C# and .NET are underrated foundations for agentic software development, citing build- and run-time type checks, Roslyn analyzers and source generators, EF Core schema checks, and a mature toolchain as guardrails for agent-written code. It highlights Aspire as a programmable runtime orchestration layer with isolated networking for worktrees and a queryable OpenTelemetry sink, and CSharpRepl as a way for agents to connect to a running application and wrap or replace functions at runtime. The piece is Part 1 of a five-part series and links the full repo at https://github.com/zeeq-ai/zeeq-tmpl.

  18. Hacker News AI20

    Controlling your AI agents from a game handheld

    The author describes using a Retroid Pocket 5 Android gaming handheld, paired with a portable OLED screen, power bank and foldable keyboard, to control personal AI agents on a mini PC during business trips. Software includes Termius, Tailscale, pi/opencode, herdr for multiplexing, and a web-based VS Code server for editing files and viewing diffs. The setup has so far been used in hotels and airports, and the author notes most Android terminal apps also support Mosh.

  19. NVIDIA Newsroom22

    How AI Helps Recyclers Turn Mixed Waste Into Reusable Material

    NVIDIA describes how four Inception startups apply AI and robotics to waste sorting: AIRECO uses vision AI and Jetson Orin NX modules to finely sort plastics, EverestLabs AI deploys computer vision-guided robots on recycling lines, Greyparrot's Analyzer platform identifies objects on conveyor belts, and reverse.fashion automates textile sorting with high-resolution computer vision. The article cites figures including less than 7% of materials entering the economy coming from recycled or reused sources and over 2 billion tons of municipal solid waste generated per year. It also notes company-reported results such as AIRECO's 95% recognition accuracy on its test line and EverestLabs' estimate of about 25,000 tons of material recycled.

  20. Hacker News AI42

    Show HN: Daily GPU Rental Prices Across AWS, Azure, RunPod and Vast.ai

    This page lists per-GPU-hour rental prices in US dollars for 11 November 2026 across AWS, Azure, RunPod Secure Cloud, RunPod Community Cloud and the Vast.ai marketplace, covering models from B300 and B200 down to RTX 4090. Prices are computed by dividing a machine's hourly list price by its GPU count, exclude storage, network and support, and omit committed-use or reserved rates; Vast.ai figures are the median and lowest of that day's on-demand offers. Daily price history runs since 02 November 2026, and list prices moving more than 50% in a day are held at the previous value until checked.

  21. Hacker News AI63

    Show HN: A company running in an AI harness

    The author presents Restless, a system for running a small "company" of AI agents inside a persistent Linux machine, where human users set direction and agents handle routine work such as research, builds, QA and scheduling. The post walks through an illustrative day at a fictional game studio, describing roles for people and agents, approval gates for external actions, model choice per agent, and pricing tiers. It also cites FrankenSurf, an open-source project with 100 GitHub stars, as a real example of a Restless company writing its code, shipping releases and marketing it.

  22. AWS Machine Learning Blog58

    Introducing Kimi K3.1 Pro on Amazon Bedrock

    Moonshot AI's Kimi K3.1 Pro Highspeed is now available on Amazon Bedrock, described as a post-trained version of Kimi K3 with the same 3T-class open-weight architecture and 1-million-token context window, trained with more post-training compute for long-horizon agentic work and fewer factual errors. It generates output tokens at a higher rate, with Moonshot AI reporting 2–3 times the speed of Kimi K3.1 Pro, and uses the same endpoint, APIs, and features as Kimi K3, so existing code can switch by changing the model ID. The post covers getting started via the Chat Completions, Responses, and Converse APIs, lowering costs with the Flex tier, explicit prompt caching, and using the model in OpenCode.

  23. OpenRouter Announcements67

    MCP vs CLI vs Skills: What Each Costs in Context

    The article measures input-token costs for giving an agent GitHub access via the GitHub MCP server, the gh CLI, or a gh skill across Claude Haiku 5.5, Claude Sonnet 5.5, GPT-6 Luna, and Gemini 3.8 Flash; on Sonnet 5.5 the CLI used 1,421 tokens, the skill 3,315, and the MCP server 59,417 with its 46 default tools. It reports that the GitHub MCP server's 46 default tools add 19,436 tokens to every request on Sonnet 5.5, and that a single tool result (GitHub's release object) cost about 20,400 tokens. It also describes the beta Tool Search server tool, which cut the up-front cost of 46 tools to under 600 tokens and halved whole-task cost on Sonnet 5.5 while raising it on Haiku 5.5.

    Why it matters: It offers concrete, model-specific token and cost measurements for choosing between MCP, CLI, and skills, plus practical mitigations like scoping tools and trimming results.

  24. OpenRouter Announcements58

    LLM Quantization: What fp8 and fp4 Do to Quality and Price

    OpenRouter tested Qwen3.8 27B across six endpoints at fp16, fp8 and fp4 on 240 exact-answer tasks and found no detectable accuracy difference, with all endpoints scoring between 98.3% and 99.2%. Across 1,455 endpoints, fp8 was the most common labeled precision (23.6%) while 59.5% reported none, and lower precision was usually but not always cheaper. The post explains fp8 and fp4 formats and shows how to pin precision with provider.quantizations, noting that a narrow filter can leave a single endpoint exposed to rate limits.

  25. OpenRouter Announcements58

    LLM API Rate Limits: How to Fix OpenAI and Claude 429 Errors

    A guide to diagnosing 429 Too Many Requests errors from OpenAI, Anthropic and OpenRouter by reading rate-limit headers and limit_source, with tested fixes: retrying at the server-provided reset time, routing overflow to a fallback model, and letting another provider serve the same model. In the author's runs, default OpenAI SDK retries left 23 of 60 requests failed, while reset-aware retries completed all 60 in 89–92 seconds (Python) and 106–218 seconds (TypeScript), and a models fallback finished all 60 in 13 seconds for $0.000066. It also covers provider failover, which turned runs with up to 32 failures per 150 requests into 150 successes, plus habits for avoiding limits.

  26. AWS Machine Learning Blog40

    Build a daily dependency watcher with LangGraph and Amazon Bedrock AgentCore web search

    The Web Search API in Amazon Bedrock AgentCore is now available, returning current web results with title, URL, publication date and query-relevant extract, signed with existing AWS credentials and served from an Amazon-operated index. The post walks through building a LangGraph dependency watch that researches npm, PyPI and container base images in parallel, keeping the date window, domain blocklist and search budget in code and checking each finding against its cited sources. In the test run, three branches made 12 searches and finished in 36 seconds versus 99 seconds sequentially, and the watch can be hosted on AgentCore runtime or offered through AgentCore Gateway.

  27. PyTorch Blog71

    ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser

    ExecuTorch's new WebGPU backend runs PyTorch models exported with torch.export to a .pte file directly on the GPU in WebGPU-enabled browsers, without converting to another framework's format. The same artifact runs across Chrome and Safari on Metal, Vulkan and D3D12, and on an Apple M1 Pro in Chrome it was 4–7× faster than ExecuTorch's Wasm (XNNPACK) CPU path on prefill and 2–10× faster on decode. Validated models include Llama 3.2, Qwen3, Gemma 3, SmolVLM, Depth Anything V2, SAM 2, Whisper and on-device LoRA fine-tuning, all running locally in the tab.

    Why it matters: It documents a concrete path for running PyTorch models on browser GPUs with benchmark numbers and a list of validated models, useful for developers weighing in-browser inference.

  28. Cursor Blog71

    Cursor Projects is now multiplayer

    Cursor announced that Projects, its multi-agent workspace introduced in September, now supports multiplayer so teammates can plan, build and ship together alongside their agents. Projects adds a shared context system holding requirements docs, diagrams, tech specs and mockups, plus a shared task board showing who and which agents are working on what and the PR behind each task. The feature also enables agent-to-agent communication across teammates, and is rolling out in beta to all Cursor users starting today.

    Why it matters: Useful for teams evaluating collaborative AI coding tools, as it describes a shared-context multi-agent workspace and its beta availability.

  29. AWS Machine Learning Blog40

    The agent evaluation playbook: Built-in, DeepEval, and AutoEval on Amazon Bedrock AgentCore Evaluations

    This AWS post explains how to evaluate production agents on Amazon Bedrock AgentCore by starting from business outcomes, breaking them into observable behaviors, and mapping each to a trace- or session-level evaluator. It covers the built-in AWS evaluators, managed third-party metrics from DeepEval and AutoEval, custom and multi-dimension LLM-as-a-judge evaluators, and suggested evaluator stacks for customer support, research, and domain-content agents. It also describes three evaluation modes—on-demand, online, and batch—that share the same evaluator IDs, plus the ListEvaluators API for retrieving the full catalog.

  30. Pydantic AI Articles38

    Turn down OpenTelemetry noise without a redeploy

    Pydantic Logfire's Fleet Management lets you connect a service and its host Collector via OpAMP and change log level, attribute redaction and trace sampling while both keep running. The post walks through a FastAPI app with fleet=True, publishing SDK policies to drop debug logs, mask *token* attributes and sample 10% of traces, plus installing a host Collector and adding nginx logs. It also covers early-access rollouts, Kubernetes and existing Collectors, and troubleshooting.

  31. Pydantic AI Articles60

    Code mode is not a sandbox

    This post distinguishes code mode, where the model writes a short Python program calling your tools as functions via a single run_code tool, from a sandbox, which is an isolated environment where code runs. It describes three places a sandbox boundary can go (the whole agent, only the tools, or only the model-written code) and how Pydantic AI workspaces let you move that boundary without rewriting tools. It also covers Monty, a Rust-written Python interpreter that runs run_code next to the agent with no filesystem, network, or environment variables.

    Why it matters: Useful for teams deciding where to place isolation boundaries when combining code mode with sandboxed tool execution.

  32. Hacker News AI60

    AgentGuard – Transaction guardrails for AI agents that hold wallets

    AgentGuard is an open-source (Apache 2.0) transaction guardrail layer for AI agents holding crypto wallets: before signing or broadcasting, an agent submits the proposed transaction to AgentGuard, which simulates it against live chain state, evaluates the effects against an operator-defined policy, and returns allow, block or escalate with an explanation, a 0–100 risk score and an immutable audit record. It supports EVM chains and Solana, decodes common drain vectors (unlimited approvals, setApprovalForAll, permits, zero-minimum-output swaps, SetAuthority, never-seen contracts), and ships a conservative default policy plus a TypeScript SDK that wraps any signer so signTransaction/sendTransaction are checked transparently. An optional separately licensed @agentguard/behavior package adds per-agent behavioral anomaly scoring, and both it and the reputation layer fail closed on outages or unreadable profiles.

  33. Hacker News AI42

    Show HN: Notebook AI – local-first meeting notes for Mac

    Notebook AI is a desktop-first AI note taker built as a monorepo with an Electron/React/Vite desktop app, a Next.js web app, a local FastAPI/LangChain sidecar, and a hosted FastAPI cloud API, released under the MIT License. Notes are stored as markdown on disk via the local sidecar, and desktop notes, local provider keys, transcription, and AI work without signing in, while Google Drive mirror and the web account vault require an account. The project lists local and cloud API endpoints, setup guides, and macOS release build commands.

  34. Hacker News AI58

    An AI agent said the code was verified. The evidence said otherwise

    A blog post argues that when an AI agent implements, tests, reviews and reports on its own code, its claim of "verified" is untrustworthy, and proposes binding evidence (review, tests, human approval) to a specific Git tree hash so that changing the tree invalidates the evidence. It introduces Uncle (github.com/unclehq/uncle), an open-source tool that runs checks itself, records exit codes and tree identities, and distinguishes "verified," "signed" and "human looked at it" rather than blurring them. The author notes the tool does not establish correctness, is young, and is partly defeated by a hostile agent with unsandboxed shell access.

  35. Hacker News AI55

    Apple has new Mac app coming to make running local AI models easy

    Leaker Aaron Perris found evidence of a new Apple Mac app called Harbor that turns a Mac into a private AI inference server for running open-weight models locally on Apple silicon. Harbor can connect multiple Thunderbolt 5 Macs to pool unified memory for larger models and higher concurrency, and is currently in a private preview under Apple's Claris brand.

  36. n8n Blog22

    Batch processing patterns for workflow automation at scale

    This guide explains how batch processing in workflow orchestration differs from stream processing and presents three patterns—chunking, rate limiting, and per-item error handling with retry—for handling large loads without hitting memory limits or API quotas. It describes how n8n's Split Out, Aggregate, Loop Over Items, HTTP Request, Wait, and Remove Duplicates nodes support these patterns, and notes n8n was tested with up to 220 workflow executions per second on a single c5a.large instance running a two-node workflow. It also covers idempotency, checkpointing, and choosing batch sizes, plus sub-workflows and queue mode for scaling.

  37. n8n Blog15

    Microservices best practices every automation engineer should know

    This article outlines microservices best practices, including defining clear service boundaries via domain-driven design and the single responsibility principle, and building resilience into service-to-service calls with circuit breakers, retries with exponential backoff, and bulkheads. It also covers routing traffic through an API gateway and making failures observable with distributed tracing and centralized logging, noting how n8n workflows can orchestrate and monitor service calls.

  38. Latent Space50

    Designing for Agents — Zeno Rocha, Resend

    This podcast episode features Resend co-founder and CEO Zeno Rocha discussing Agent Experience Optimization, including an AEO study in which Claude Code picked Resend first in 64 of 102 email recommendations, and how React Email templates and a pricing page readable via a .md URL helped Resend reach developers before models. Rocha describes growing Resend's MCP server from roughly 20 tools at launch to more than 80, weighing that against Cloudflare's two-tool Code Mode server, and explains why Resend introduced custom tracking domains to keep one customer's blocklisting from affecting others. He says Resend had 4.3 million users and 120,000 paying customers at recording time.

  39. Mastra Blog31

    $25 Starter Plan on Mastra Platform

    Mastra introduced a $25 Starter plan that raises monthly allowances for compute, data, and observability events and includes one Small dedicated server whose compute is covered by the $25, with usage beyond the allowances billed on top. Compute measurement is changing and allowances are being updated on every plan, while a serverless agent remains free to run on all plans including Free. Projects run serverless or on Small, Medium, or Large servers, and Mastra Connect now has a connection allowance on each plan (5 on Free, 10 on Starter, 50 on Team, $0.50 each additional) with egress from those connections counting toward plan egress usage.

  40. Hacker News AI45

    Doc2Harness: Benchmarking how well AI builds agent harnesses

    The authors propose Doc2Harness, a two-stage benchmark that evaluates long-horizon harness engineering: a build model constructs a candidate harness from an interface-only contract skeleton and a development document consolidating behavior across twelve frontier harnesses, and the frozen harness is then run per test case by a scripted mock model and scored via four observable channels. Across nineteen frontier models, each with three independent builds on 328 test cases, GPT-6 Astra achieves the highest pass rate at 53.05% while the weakest model reaches 27.44%, and failures cluster around how cross-cutting functionality is split across modules. The authors describe it as the first deterministic benchmark for measuring long-horizon harness engineering.