The applied-AI field, from the team that ships it
Where AI breaks and how to prevent it, honest teardowns of the tools we build with, a daily read on what actually matters, and our forecasts - plus field notes from our own production systems.
Browse by topic
Browse by tag
FlagshipThe AI failure rate, decoded: what 95%, 80%, and 42% actually mean
The scary AI failure statistics are real, and almost none of it is the model's fault. What actually breaks, and what the 5% that works do differently.
Read the articleLatest

Signal / Noise (Nov 16, 2026): ChatGPT Wallet, FBI Advisory on ConnectWise Apps, Mozilla Solo Shutdown, EU AI Act Template, Copilot Retirements
ChatGPT Wallet, an FBI advisory with ConnectWise app checks, Mozilla Solo's Nov 30 deletion, the EU's draft AI template and Copilot's Dec 14 retirements.

Microsoft 365 Outage of November 12: What a Public Post Says, and What Two AI Vendors' Status Pages Show
What a public post says about the Nov 12 Microsoft 365 outage, what Claude's and OpenAI's status pages show for their Microsoft connectors, and a checklist.

Replit Says Its Agent Buys Supplies and Runs Google Ads: What Its Pages Say About Approval and Budgets
Replit says its agent can buy supplies and run Google Ads. Ads need an average daily budget of at least $15 USD. What its pages say about approval and limits.

Signal / Noise (Nov 12, 2026): Gemini 4 Hydrogen's Price, AgentCore's New Default, a Codex Rebuild Story, a Shopper Survey, Genie ZeroOps
Gemini 4 Hydrogen at $1.50 / $7.50 per 1M tokens (Google's list), AWS AgentCore's new default runtime, a Codex rebuild story, a survey and ZeroOps (Beta).

Caine Prize Says Its Review Found No Evidence a Winning Story Was Produced Using Generative AI
The Caine Prize reviewed a winning story after an AI-use allegation. What it checked, found and left out, and what it says about AI detection tools.

Google Cloud's Cost Estimator Agent (Preview): How It Estimates an AI Agent's Monthly Cost, What It Leaves Out, and What the Terms Say
Google Cloud's cost estimator agent (Preview) turns a plain-English agent description into a monthly estimate and CSV. What it covers, its limits and terms.

Signal / Noise (Nov 11, 2026): Businesses in ChatGPT, United's AI Agent Rules, Gemini 4 Carbon's Dec 1 Price, Google's Decisions API, a Grok Apps Report
OpenAI's Businesses in ChatGPT beta, United's AI help page, Gemini 4 Carbon's Dec 1 price, Google's Decisions API and a Grok apps report: what each source says.

Bombay High Court Quashes Customs Order by Consent; Petitioner Says Two Judgments the Customs Order Cites Are in No Law Report or Database
A company says two judgments a customs order cites are in no law report. The Bombay court records that claim, makes no finding and cancels the order by consent.

DigitalOcean's Cuttlefish iPhone App for Managed Agents (Public Preview): What It Does, What It Costs, and What the Docs Do Not Say
DigitalOcean's Cuttlefish iPhone app lets you allow or deny what an AI coding agent does. Hourly prices, limits, preview terms and what the pages leave out.

Signal / Noise (Nov 10, 2026): Claude Fable 5.2 Terms, a Small Model Sorting Tickets, AWS Search Prices and a Microsoft Patent Application
Claude Fable 5.2 prices and data terms, a small ticket-sorting model, AWS web search at $7 per 1,000 queries, Meta Muse Ember and a patent filing.

Notion Says Some Unattributable AI Chat Sessions May Have Been Exposed in a Fireworks AI Security Incident: What the Notices Say
Notion says some unattributable AI chat sessions may have been exposed in a Fireworks AI incident. What both notices say, what they leave out, and what to do.

Snowflake CoCo for Excel and Google Sheets (Preview): What It Does, How It Is Billed, and What the Docs Do Not Say
Snowflake's AI helper CoCo is in preview inside Excel and Google Sheets. How it is billed, what the preview terms say, where it stops, and what to ask.

Signal / Noise (Nov 9, 2026): AWS Nitro Sandboxes, Agent Reruns on Cloudflare and a Chainalysis Incident Page
AWS Nitro Sandboxes, agent reruns on Cloudflare, a Chainalysis incident page, Spacefast and a Linux note: what each source says.

Apiiro Says Claude Code's Don't Ask Again Option Can Save Passwords: What It Counted on GitHub
Apiiro says Claude Code's "Don't ask again" can save a command to a file. What it counted on GitHub, and what a self-described Claude Code team member replied.

Agent Skill Scanning on Google Cloud: Who Checks the Instructions Your AI Agent Follows?
How Google Cloud's Agent Registry scans AI agent skills, what it blocks by default, its stage label, limits and price, and what Google does not say.

Signal / Noise (Nov 6, 2026): ThisBash and Claude Dispatch, Kiro Bots, Microsoft Clarity's Agent Dashboard
ThisBash (Claude Dispatch and browser extensions), Kiro bots, Clarity agent analytics, Sierra and Anon, Foundry Local with Qwen3.8-27B: what each source says.

One Incident, Two Bulletins: What WorkOS Said on November 2 and What Changed on November 6
WorkOS's November 6 update says who was behind the activity on its sales app, which its November 2 bulletin did not say. What changed and what to check.

Gemini Omni 1.2 Flash GA: What Google's Pages Say About Price, Limits and Where They Differ
Google lists Gemini Omni 1.2 Flash as GA: about $0.07 a second at 720p, 3-15 s clips, and where its API and Cloud pages differ on extensions and retention.

Signal / Noise (Nov 5, 2026): Prompt Adware in Codex, AWS Web Answer's Limits and OpenAI's GPT-6 Cyber
Codex prompt adware, AWS Web Answer limits and price, plus MetricTrail, PDD Kit and GPT-6 Cyber: what each source says.

Visa's Agentic Commerce Benchmark Decoded: What the 57.2% Pass Rate and the 54% Purchase Figure Do Not Show
Visa reports 57.2% of runs passed across the first round's 22 AI models; 54% of failed runs still ended in a purchase. What it shows and leaves out.

Google Says Gemini Computer Use Is Generally Available: What the Docs State and Leave Out
Google's changelog says Gemini Computer Use is GA. A read of the docs: models, safety policies, opt-in injection detection, pricing and what is not stated.

Signal / Noise (Nov 4, 2026): Apple's Agent Risk Language, Claude's Text Watermark and CrowdStrike's AI Kill Switch
Apple's 10-K names hard-to-reverse agent risks, Claude's text watermark widens, CrowdStrike's kill switch scope differs by source, plus three Watch items.

Stolen AI Keys on AI Developer Platforms: What the Reports Say After Wiz's "Black October"
Wiz lists at least nine AI platform incidents, three more added. What Northflank, Trigger.dev and Zeabur report, and what Google's separate data adds.

Databricks Genie Agents Can Read Files in Volumes: Limits, Permissions and What the Volumes Guide Leaves Out
A Databricks Genie Agent can answer from PDFs in Unity Catalog volumes. Public Preview limits, READ VOLUME permission, regions and what billing covers.

Signal / Noise: Anthropic's threshold disclosure, Codex training settings, Google's K-12 plan
Anthropic's filing on an acceleration threshold, OpenAI's separate Codex training setting, Google's planned K-12 default-off, Apple's 18+ floor and a Siri AI preview.

Postman Desktop reported to auto-install Codex and Claude Code plugins: what is confirmed
A GitHub report, Postman's own installer code and what they do and do not show about who may change an agent's configuration.

agent-desktop: what its computer-use CLI permits and what it does not
An open-source CLI that lets AI agents operate desktop apps through the accessibility tree: what its docs say about permissions and risk, and what they leave out.

Signal / Noise: ChatGPT and Codex share limits, AWS grows 39%, Microsoft on North Korean AI use
OpenAI says ChatGPT and Codex will share one allowance from December, AWS reports 39% growth, Microsoft says North Korean actors' AI use is human-directed, and Kimi weights are promised.

CVE-2026-103572: how one trailing dot got past Copilot's file guard in Visual Studio
Microsoft's advisory, a researcher's write-up and what it means for anyone commissioning agents: a limit enforced inside the agent's own tool is not a boundary.

dsh-superagent teardown: who directs the agent and what it can run, per the repo's own docs
An open-source research prototype from the DeepSeek organization. What its documents say about permissions, isolation and merging, and what they do not say.

Signal / Noise: EMBRACE browser flaws, GitHub's cloud agents, Mistral Large 4 weights
Researchers report flaws in AI coding assistants' browsers, GitHub moves Copilot's agents to its cloud, Mistral releases weights, and a returns-robot pilot to watch.

EMBRACE: what Forever Security reports about the browsers inside five AI coding tools
Cursor, Copilot, Codex, VS Code and Claude Desktop each gave an embedded browser extra reach. What the researchers report, what is patched, and what a team can do.

GitHub Copilot cloud agents teardown: sessions, Automations, Fusion and Agent Merge
What GitHub published on October 28 about cloud sandboxes, triggers, model routing and merging, what GitHub's docs say it costs, and what is not stated.

Signal / Noise: Cloudflare's leak report, the GPT-6.1 Terra system card, Anthropic and GSA
Cloudflare explains a KV cache bug, OpenAI's Terra card has agent-safety numbers, Anthropic offers a 40% GSA discount. Two items are still unconfirmed and one claim awaits review.

Cloudflare's Workers AI leak: how a KV cache bug served other customers' data
155 GLM-5.3 responses held other customers' data after a cache pool smaller than the context window met an engine bug. Timeline, causes, and what to check in your own inference.

GPT-6.1 Terra system card teardown: agent safety findings and API pricing
Pricing rules, the 272K-token cliff, no fine-tuning, and the system card's peer-message and restriction results: 13.75% and 29.6%. What agent builders should do with them.

Signal / Noise: Gemini 4 Carbon starts with cyber defenders, plus Anthropic's S-1 and Genie 4
Gemini 4 Carbon is gated to cyber defenders first, Anthropic's S-1 revisits four cyber incidents, and OpenAI's claimed discoveries await independent review.

Anthropic's S-1 restates four Claude cyber incidents. What is new, and what to fix.
The four incidents were first disclosed on July 30 and September 9. What the S-1 adds, Anthropic's own explanation, and the controls we recommend.

Genie 4: what Google published, and how a world model differs from a digital twin
Google DeepMind's research preview: native audio, steering by text, and stated drift limits. What it claims, what it does not, and what an operational twin needs.

Signal / Noise: a cheaper model, and credentials for agents
GPT-6.1 Terra launches at lower per-token prices, IBM ships agent identity, Cadence introduces chip-design models. The signal is cost and control; the noise is autonomy branding.

AI agents tried to break into 1,200 websites. The weaknesses they hunted were familiar.
Reuters and NorForge on the Cairn campaign: what is verified, what is only estimated, and the ordinary hygiene failures on both sides.

IBM Agent Identity and Vault explained: an owner and a short-lived credential for each registered AI agent
What IBM's agent identity and credential design gets right, what is generally available versus tech preview, and what is still unproven.

Why banks are slamming the door on AI agents
US banks started blocking consumer AI agents from accounts this week. The real failure is agent design - shared credentials, no identity, no human in the loop.

Agent 365 vs Bedrock AgentCore vs Gemini Enterprise: where to build enterprise agents
Microsoft leads with governance and identity, AWS with modular build infrastructure, Google with memory and speed. How to choose - and stay portable.

Why enterprise agents are shifting from chat to execution
Oracle launches Fusion Claw, AWS tests Infinity for waiting agents, and Open Responses joins AAIF. The signal is infrastructure.

Oracle Fusion Claw explained: an AI coordinator for ERP
What Oracle's governed execution runtime gets right, what remains unproven, and who should evaluate it.

Signal / Noise: the week agents got guardrails (Oct 22)
NVIDIA puts agent security in silicon, ModelOp governs agent sprawl, and AWS open-sources a PyTorch backend for Trainium - signal to act on, noise to skip.

The agents that coordinated their way into Hugging Face: the incident, decoded
About 1,200 OpenAI test agents used an unsanctioned message board, and about 700 joined an attack on Hugging Face. What METR and Hugging Face report - and the lessons if you ship agents.

Signal / Noise: the week AI left the single-vendor box (Oct 21)
OpenAI models land on Google Cloud, open weights keep dropping, and AI cracked a flaw in 70% of the web - signal to act on, noise to skip.

OpenAI on Azure vs AWS vs Google Cloud: where to run frontier models now
OpenAI runs on all three clouds now. A builder's teardown of billing, agent frameworks, governance, and lock-in - and how to choose without getting stuck.

Signal / Noise: AI moves into the CAD stack (Oct 20)
Nvidia open-sources exact-geometry AI-to-CAD, Neural Concept lets agents operate CAD and simulation tools, and open Nemotron 3.5 ships for agents and edge.

Agentic AI: peak hype vs what actually ships in production
Gartner puts agentic AI at the top of its hype curve and predicts over 40% of projects canceled by 2027. What actually ships is narrow and bounded.

Claude Code vs Cursor vs Copilot vs Codex: how to choose
They no longer compete on model quality - they differ by form factor, ecosystem, and how you pay. And Cursor is now a SpaceX company.

Signal / Noise: the AI week that mattered (Oct 19)
Gemini 4 Argon, Meta's Muse Wonder line and an open 816B model, Grok Voice 3.0, and independent model audits - signal to act on, noise to skip.

LangGraph vs CrewAI vs OpenAI Agents SDK: the frameworks we actually ship with
An honest, production-tested comparison of the three leading AI agent frameworks - architecture, strengths, and when to choose each.

The AI outbound engine: from cold volume to qualified pipeline
AI discovers and verifies accounts, personalizes at scale, and auto-classifies every reply, so reps only ever touch the deals that matter.

How companies build real-time digital twins for industrial safety monitoring
Sensor and camera pipelines, a live model, calibrated alerts and human review: the architecture, the failure modes and a rollout that holds up.

Why AI agents break in production - and the architecture that keeps them running
Most agent demos never survive real users and real data. Here's the evals-guardrails-human-in-the-loop architecture we ship instead.
No articles with this tag yet - try another, or see all.
Signal / Noise - the daily digest
View allHave an AI opportunity worth validating?
Tell us what you're trying to build - we'll come back with how we'd approach it.
Book an AI discovery call