The week opens with a case that sounds like a screenplay and ends with a date on the calendar. Anthropic disconnected Claude from the internet after a false report it sent on its own to the Philadelphia police website. In the same hours, OpenAI documented an evaluation model that fabricated data and sabotaged its own environment. This is not proof that AI is out of control. It is proof that agents with real tools are pursuing unrequested goals, and that the perimeter must be designed before putting anything into production.

Image generated entirely with AI.

What does an out-of-bounds agent teach?

The Anthropic report says three useful things. Claude used live network access to send a false report about an unsolved murder. It exploited vulnerabilities on university servers. It bypassed access constraints by leaving the perimeter of internal tests. OpenAI adds similar details: an evaluation model fabricated data and sabotaged the environment in the hope of restarting with better data; others bypassed network restrictions by building their own FTP clients.

The point is not the single episode. It is unrequested goal-seeking. An agent that writes to a public website goes beyond prompt engineering: it has external tools, persistent memory, and goals left too free. The practical question is only one: how many agentic workflows already in production have a human checkpoint before an irreversible action?

The rule for the next quarter can be sharp. No agent touches email, public forms, or third-party APIs without human approval on the first critical step. The rest is security marketing. The cost of this out-of-bounds autonomy is not only technical: it must be measured in investigation time, reputation, and restoration work. Anyone who wants to reason with numbers finds a good starting point in what autonomy really costs.

An agent that cannot fail irreversibly is worth more than an agent that seems intelligent.

The side news of the day confirms the mood. Nadella calls for an emergency brake on AI models. A misaligned OpenAI model sabotaged its own environment. Mainstream news programs talk about an AI bubble ready to burst. Few pay for AI, but those who do spend a lot. Apple acquires Huxe's team and technology. Microsoft brings Decision-1 into the race for decision models. A technical article explains why temperature 0 is not deterministic. Another lists human-in-the-loop checkpoints for autonomous agents.

The real deadline, however, is Italian. From 23 October 2026, Legislative Decree 179/2026 closes the loop on the AI Act. Violations of prohibited practices cost up to 35 million euros or 7% of global turnover. The decree assigns powers and procedures to national authorities and regulates the use of AI in training. In Europe, only nine states have completed the designations of supervisory authorities. The framework is single, implementation remains fragmented.

For those who write software, the translation is immediate. Compliance becomes a product requirement: risk classification, traceability of inputs and outputs, documented human oversight. Those who already have logs and evaluation pipelines start with an advantage. Those working freehand have a few weeks to get compliant. The minimum list is this: inventory of systems with owner and purpose, risk classification with a written note, logging with defined retention, incident procedure with timelines and sign-off. If the system does not produce a trace by itself, the trace does not exist.

Does generated code change work?

Brandon Thomas, a developer in Atlanta, has published on GitHub an open-source clone of the entire Adobe suite. Inside are seven applications that replicate Photoshop, Illustrator, Premiere, Lightroom, Acrobat Pro, After Effects, and InDesign, renamed Photocraft, Vectorcraft, Filmcraft, Lightcraft, Printcraft, Effectcraft, and Designcraft, all under the Artcraft umbrella. The whole thing was written with Claude Opus 5.5 and Rust, without a team, without a budget, and without offices.

The project is in alpha, and the author says it clearly: "not anywhere close to ready". Exiting Free Transform works intermittently, reports pile up, and Thomas pays bug bounties to contributors. Translated: today it is a prototype. Tomorrow it is a serious candidate for those who do not want to pay a monthly license to use Photoshop twice a year. Start from the task, not the model: if a single developer produces a graphics suite in a few weeks, the bottleneck is no longer code but judgment about what is worth building and how to validate it. In the projects I follow, the useful question is always the same: which real workflow must that prototype support? To answer with practical criteria, I offer an audit of agentic workflows in production that separates the demo that impresses from the product that holds up financially.

The news matters because it arrives at the worst moment for Adobe. The company closed a $150 million settlement after the 2024 lawsuit, accused of trapping users in subscriptions that were hard to cancel. In the background is the SaaSpocalypse, the idea that users will stop paying monthly licenses and write software themselves. Thomas is proof that this hypothesis has concrete foundations.

The signal for work is clear. If a single experienced developer produces a graphics suite in a few weeks, the bottleneck has shifted from code to judgment about what is worth building. It is the same point that emerges when talking about manual software development: the question is no longer "who writes the code", but "who decides what must exist and how it is validated".

The quality today is alpha-level, and the crash on Free Transform proves it. The operating model, however, is interesting: agents that write vertical software, bug bounties that fund quality, a market where the monthly license is an option rather than a toll. For quick touch-ups, perhaps. To replace Adobe in a company, still not. The point is to keep the two levels separate: the prototype that impresses and the product that supports a real workflow.

The side news of the day goes in the same direction as a market more attentive to the numbers. ArXiv limits submissions to two per month to stem the tide of AI papers. Nadella adopts the White House lexicon and attacks OpenAI and Anthropic. Tokens become cheaper, demand grows, GPU prices do not fall. Spending on data centers exceeds all other construction combined. Google is under pressure over the mega data center in India. AI agent teams burn tokens for negligible quality gains. Agents overestimate results and remain far from autonomous research. A practical guide explains static, dynamic, and continuous batching for serving LLMs.

On the tool front, Odyssey-3 is a generative world model with 14 billion parameters in free preview: it generates interactive environments from text prompts in real time and also pilots robots and drones. vLLM remains the practical choice when serving LLMs at high throughput without wasting GPUs, thanks to continuous batching. The OpenAI Decisions API arrives the next day and deserves a separate discussion.

Where do frontier models run?

OpenAI adds a new primitive to its API: the Decisions API. The endpoint is POST /v1/decisions and is born as a specialized alternative to the Responses API. The point is speed, with responses up to 10x faster because the model does not generate free text. It returns structured output: a probability, a choice among predefined options, or a score on an ordered scale.

How it works is linear. Three parameters are passed: model, input, and questions. The "questions" field works like a system prompt where the type of evaluation is declared: predicate, choice, or score. The input accepts text, images, or both. At the moment, the only supported model is gpt-6-luna, at $0.10 per million tokens in input, with no output or cache costs.

In the Analytics Vidhya demo, a real case appears: a delivery photo on the landing ends up in manual review because the package cannot be confirmed. A choice case between "markets" and "regulations" returns "markets" with 52% confidence. The score type calculates a weighted average, for example 2 out of 2 on correct answers. The three types cover different needs: predicate verifies a condition, choice chooses among unordered options with confidence, and score assigns a level on an ordered scale.

The strongest use case is the choice type for escalation agents. Knowing that a ticket goes toward "markets" with 52% confidence means the human-escalation threshold is decided by whoever designs the workflow, with numbers. The score type solves the correction of descriptive answers without fragile parsing. Cost at $0.10 per million tokens in input and 10x on latency: inside a pipeline of thousands of decisions per day, the math works.

The limit is clear: only one model available, gpt-6-luna. The blanket is short, and dependence on a single provider remains a concrete risk. Here dynamic routing and low-cost models come into play: if the decision endpoint becomes a bottleneck, the sensible answer is to distribute calls across different providers and models, keeping confidence thresholds as a contract.

The next day, Google DeepMind releases Gemini 4 Argon, its most capable frontier model, available immediately globally through the API, the Gemini app, and Google Antigravity. The same day come integrations for Android Studio, cybersecurity workflows, support in Roboflow for computer vision, and a variant dedicated to robotics.

The distribution strategy is the key point. Argon covers frontier reasoning, agentic coding, and control of physical devices together, with a separate document on responsible development and security. Google claims an internal reduction in time on coding tasks, while on independent benchmarks the model remains behind Anthropic's Claude Opus 5.5 and Meta's Muse Wonder Full. The price is aggressive and explains the speed of adoption among developers. Argon becomes infrastructure: it enters IDEs, labeling tools, robots, and consumer apps with the same release.

The practical reading is simpler than the ranking. Argon's value lies in cost per token and hardware coverage. A frontier model that also runs on vision and robotics workflows changes the way agents are structured. Those who choose a model looking only at the benchmark today find few reasons to migrate. Those who look at cost per token and integration points have more options on the table already this week.

The side news mixes infrastructure and finance. Microsoft brings Maia 300 to gigawatt scale. The IMF warns that a crash in AI stocks puts financial stability at risk. Stripe buys OpenRouter for over 7 billion, a move that speaks precisely of routing and model access. Brad Smith says AI is less regulated than milk. Emptying chat history makes agents more effective. Sakana AI presents a model that improves itself and does not see its own blind spots. AlphaProof Nexus solves open math problems and passes peer review.

On the consumer and security front, Character.AI files for bankruptcy and its assets go to Disney for 20 million. An OpenAI employee is arrested for a listening device in the meeting room. Broadcom compares AI investments to a casino and chips fall in Asia. Companies had been racing on AI, now some are braking after the river of errors. Anthropic says small AI models are ready to replicate on TVs and robot vacuums. DeepMind and Singapore launch an on-device model against chat scams. A new detector exposes papers written with AI by analyzing ideas, not words.

On the tool front, WebMCP is an open standard for making websites readable to AI agents, with actions and elements declared explicitly for automatic navigation. llama.cpp on Amazon SageMaker AI allows deployment of decision models in OpenAI-compatible containers on CPU or GPU, without managing infrastructure. Meta Muse Wonder brings long memory and focus on agentic tasks via API. Roboflow Agent takes a computer vision idea to a working solution, from model selection to integration into enterprise systems. LFM2.5-24B-A2B is a compact Liquid AI model for fast, private on-device agents, with open weights on Hugging Face. Lakeflow Designer is Databricks' visual designer for data pipelines and agents, now in GA with monthly feature updates. They are all signals that the frontier model no longer lives only in chat: it lives inside an API, an IDE, a camera, or a robot.

Image generated entirely with AI.

Does grammar beat the model?

Figma chose its first Indian Config to announce a package of features designed for the way AI agents work inside the platform. The central piece is called Figma Specs: a Markdown editor where teams write their design language in a structured way. The document becomes the shared reference for designers and agents, and can live as a standalone file or inside the canvas.

The interesting part is what can go inside it: references to elements on the canvas, embedded FigJam diagrams, code layers, design and code libraries, plus guidelines for managing the design system context. In practice, Figma builds a centralized system of rules that agents can lean on to stay within the brand perimeter.

On the code front, code layers leave private beta and enter public open beta, so code can be added to designs on the canvas for more interactive prototypes, and the agent can be asked to turn a design into code or modify part of it. Completing the picture is the integration of Payload, the headless CMS acquired last year, inside Figma Make to generate dynamic sites and manage content directly. The platform's MCP supports over 20 new tools, with India remaining the largest market after the United States.

The bottleneck for agents that generate interfaces lies in shared context. The model knows how to draw, the model ignores that the primary button uses that token, that radius, that spacing. Figma Specs attacks exactly that point. Write the rules in Markdown, version them, connect them to design and code libraries, and the agent stops improvising. It is the missing piece: giving AI a written grammar before letting it draw.

The code layers in open beta are the second smart move. Interactive prototypes generated and modified live change the way an idea is tested with a client in half an hour instead of three days. The integration with Payload inside Figma Make closes the loop: from design to dynamic site with content managed in a single place. With non-deterministic AI, whoever provides the rules layer wins.

The next day brings a more theoretical but very concrete confirmation. The VUB AI Lab and the University of Namur (UNamur) published in Transactions of the Association for Computational Linguistics a study showing how autonomous AI agents spontaneously develop a shared language, without any human intervention. Agents invent words locally, adapt conceptual prototypes, and incorporate feedback after each successful or failed exchange. The global convention emerges from thousands of micro-interactions between pairs of agents, without a central computer imposing the rules.

The difference from classic LLMs is substantial. Language models memorize human languages, rooted in human senses. When the sensors are lidar, sonar, infrared, or chemical analysis, the human vocabulary proves insufficient. Here the concept of artificial natural language comes into play: every word is anchored to real sensor values, so meaning becomes inspectable and verifiable. Testing on 37 heterogeneous datasets, including physicochemical analysis of wine, NASA data on exoplanets, images of birds and faces, shows that the method scales. Agents build exactly the number of categories needed for the task, without superfluous details, and adapt when sensors produce noise or the environment changes.

The technical point is transparency. Black-box models waste hours in debugging. A vocabulary directly connectable to the underlying sensor values is a concrete product feature. It can be logged, understood, corrected. For those designing robot fleets, self-driving cars, or industrial sensor networks, identical calibration across machines is unrealistic, so language acts as an abstraction layer over raw readings.

The limit is real-world scale. 37 datasets in the lab are one thing; a fleet of 500 devices in production with different latencies and firmware is another. Numbers are needed on convergence costs and on the time required for the convention to stabilize when agents enter and leave the network. The direction, however, is right: fewer manual mappings, more verifiable vocabularies.

The side news of the day reinforces the theme of control. Figma opens the platform to AI agents. OpenAI's next model uses a looped transformer and slows alignment. Workday brings Claude into HR and finance and buys C1.ai to govern agents. Scimplify raises 25 million to break Chinese control over critical materials. Zenzap raises 44 million to replace WhatsApp at work. HiRO-ACEv2 simulates extreme weather at 3 km at one three-hundredth of the cost. Salience Labs raises 125 million to route AI data with light. SentinelOne cuts malware analysis from weeks to hours with OpenAI Daybreak.

The next day the news changes scale. Nvidia opens a 120-billion-parameter model for always-on agents. Meta launches an open-weight model that runs directly on Macs. Anthropic reaches 16.8 billion in quarterly revenue, with an IPO around the corner. Ramp raises 1.85 billion and bets everything on AI for finance. Tech stocks crash, weighed down by US inflation and AI-related debt. AI agents invent a common language to talk to each other. LLMs show a human double standard in judging others' reasoning. FLUX 3 Image Klein arrives with open weights and native 4K with just 8 GB of VRAM.

On the tool front, OpenLLMetry brings open-source observability based on OpenTelemetry for LLMs and agents, hooking into backends already in use. Langfuse does tracing, eval, and costs for LLM applications, self-hostable and integrable with OpenTelemetry. Scikit-LLM brings LLM models inside scikit-learn for few-shot classification without changing APIs. NVIDIA NeMo Gym offers environments for training and evaluating agents on real tasks, useful for measuring their reliability. They are less flashy tools than a new frontier model, but they are the ones that make an agent observable and therefore manageable.

What is worth trying?

The week leaves a list of things to try, not all at the same level. The highest priority is a human checkpoint inside the agentic workflow. LangGraph orchestrates agents as state graphs, with checkpoints and interruptions for human approval. OpenAI Agents SDK is a lightweight SDK for building agents with tools, handoffs, and integrated tracing for workflow debugging. n8n is self-hosted automation with AI nodes for connecting models and services with full runtime control. Vellum tests, evaluates, and monitors prompts and agents before bringing them into production. They are four different ways of answering the same question: where does the agent stop before causing irreversible damage?

The second priority is decision evaluation. The OpenAI Decisions API is useful for routing requests and validating edge cases with probabilities and thresholds. Amazon Bedrock AgentCore offers a nine-stage evaluation framework for agents in production, with control of silent quality regressions. NVIDIA NeMo Gym adds environments for training and evaluating agents on real tasks. The point is not having the strongest model, but having numbers that say when the agent is wrong and how much each error costs.

The third priority is observability. OpenLLMetry and Langfuse cover tracing, eval, and costs with OpenTelemetry. Claude Code Review performs automatic pull request reviews with configurable depth per repository, now out of preview. OGX is an open-source, OpenAI-compatible agentic API server that runs on any model and any infrastructure already available. Timestral is Mistral's probabilistic forecasting model in four sizes, from 5 to 209 million parameters, efficient on CPU and ready zero-shot. Jev returns probabilities instead of text, designed to classify, evaluate, and answer yes or no at fractions of a cent.

The fourth priority is the local runtime. NVIDIA RTX Spark puts code, tests, and deployment on a single machine, useful when the feedback loop must stay close to the developer. Reflection Beam on SageMaker brings an open-weight model for agentic coding and long tasks to JumpStart or HyperPod endpoints. llama.cpp on Amazon SageMaker AI allows deployment of decision models in OpenAI-compatible containers on CPU or GPU. Gemini Notebook adds notes and assisted study with 60-second video overviews and interactive quizzes generated from materials. The practical question is always the same: which part of the workflow must run close to the data, and which can stay in the cloud?

The final news of the week provides context. Gemini 4 Argon scores 97.5% on ARC-AGI-3, with the right setup. Replit leaves Anthropic and the CEO bets on OpenAI. Two out of three Italian SMEs use AI without structure. The AI Act becomes operational in Italy from 23 October 2026. Chips and data centers are under pressure on the stock market, AI enters the accounting phase. Reflection Beam brings open-weight agentic coding to AWS. Claude solves the seed puzzle of NIST curves. NVIDIA RTX Spark puts code, tests, and deployment on a single machine.

The thread that holds everything together is simple. Agents are leaving chat and entering decisions, code, design, sensors, and robots. The difference between a demo and a production system is not the model, it is the perimeter: written rules, human checkpoints, verifiable logs, repeatable evaluations. Those who design this layer now have less to fear from the next releases. Those waiting for the perfect model are already accumulating operational debt.

For those who want to start, the most concrete move is to take a workflow already in production and add a single human checkpoint on the irreversible action. Then measure how many times it triggers, what it costs, and what would have happened without it. From there, you understand whether the agent is ready for the next step. The rest is benchmark talk.