When Agents Stop Waiting for Instructions
"There is a fast moving operational development in AI: AI is moving from a model-selection exercise to an infrastructure decision. Most enterprises aren't ready for that. They've picked a model, maybe run a pilot, and assumed the hard part is done. It isn't. Agents operating across real systems need identity, permissions, cost controls and someone accountable when something goes wrong. The outputs also need to be deterministic. Typed, bounded, auditable, not probabilistic text that has to be interpreted. The organisations that lead won't be the ones with the best model. They'll be the ones with a governed operating environment that lets agents act productively without losing control. This issue maps where that transition is already underway."
CEO, Magure
Akhil brings 15+ years across growth, strategy, and transformation, and has led AI adoption programs for enterprises across the Middle East and global markets, with a focus on making AI accountable, observable, and cost-resilient at enterprise scale.
State of AI Play: In Figures
- of organisations are now building, deploying or developing AI agents - up from 53% one quarter earlier
- 62%of organisations are now building, deploying or developing AI agents - up from 53% one quarter earlierKPMG
- Additional NVIDIA GPUs AWS and NVIDIA plan to deploy in 2027-2028 for agentic and physical AI
- 2MAdditional NVIDIA GPUs AWS and NVIDIA plan to deploy in 2027-2028 for agentic and physical AINVIDIA
- Share of Anthropic's internal AI R&D work that Claude led in August
- 26%Share of Anthropic's internal AI R&D work that Claude led in AugustReuters
- AI agents Huawei forecasts could be active globally by 2035
- 900BAI agents Huawei forecasts could be active globally by 2035Reuters
- Agent-workdays of effort per human workday inside OpenAI's research organization as of mid-August 2026
- 3.1xAgent-workdays of effort per human workday inside OpenAI's research organization as of mid-August 2026OpenAI
- report slowing AI deployment because of sovereignty concerns
- 8%report slowing AI deployment because of sovereignty concernsGlobal AI Pulse Q3 2026
The AI Deployment-Production Gap
| Adoption and Scale | Economics and Control |
|---|---|
| 38% of organisations are developing or implementing multi-agent systems. | Only 12% consistently measure AI-generated value against its cost across the organisation. |
| 34% report significant employee adoption of AI agents, up from 25% in Q1 2026. | Only 23% have model routing within their AI management layer. |
| 55% operate a formal AI management layer between models and business use cases. | Only 21% have an enterprise-wide AI strategy that is regularly reviewed. |
| Planned AI investment averages $210M over the next 12 months, up 13% from Q1. | Only 53% assign accountability for AI-informed decisions to the C-suite or higher. |
Source: KPMG International, Global AI Pulse Q3 2026
Global AI Moves
TypeSafe AI Launched Jev, a Decision Model for Agents That Does Not Generate Text
TypeSafe AI introduced Jev, an early-access model backed by a $40M seed from DCVC. Built by former OpenAI researcher Diogo Almeida, it converts unstructured workflow state into typed probabilistic decisions rather than natural-language output. Within three days of launch, Jev became the fastest-adopted model in Vercel AI Gateway history, with LangChain, Cloudflare, and Netlify shipping integrations in the first week.
Why it matters
Agent systems make many small decisions that typically require full LLM calls. Jev handles them at a fraction of the latency and cost - shifting agent control logic into non-generative models that enterprises will still need to test, monitor and govern.
AWS and NVIDIA Commit 2 Million More GPUs to Agentic and Physical AI
AWS and NVIDIA announced plans to deploy 2 million additional NVIDIA Blackwell Ultra, Rubin and Rubin Ultra GPUs across AWS infrastructure in 2027 and 2028. The companies said the capacity will support agentic AI, scientific discovery, enterprise automation and physical AI, including robotics.
Why it matters
The next AI bottleneck is not model access. It is the amount of compute available to run agents continuously - planning, calling tools, checking results and retrying when workflows fail.
OpenAI Released GPT-6 Astra and Launched the Agents API - A New Model and Infrastructure
OpenAI released GPT-6 Astra on September 3, focused on cybersecurity and computer automation, priced at $10 per million input tokens. On September 10, OpenAI introduced the Agents API in public beta - a managed service for building cloud agents that operate across multiple steps with long-running sessions and tool use.
Why it matters
GPT-6 Astra sets a new ceiling on what premium model access costs - and the Agents API signals where OpenAI wants to sit in your stack: not just the model, but the execution layer your agents run on. For enterprise leaders, these two moves together ask the same question: how much of the AI stack - from model inference to agent runtime - should sit with one vendor?
OpenAI Adds a Data Agent to ChatGPT Work
OpenAI introduced a Data agent for ChatGPT Work that connects company data sources, surfaces insights and creates interactive dashboards through natural-language instructions. The product is designed to work across approved enterprise data connections and files.
Why it matters
Enterprises will need to govern not only what employees can see, but also what agents can infer, combine and act across systems.
Salesforce and Anthropic Announced Claudeforce - Merges CRM and AI Into a Single Working Layer
In late August, Salesforce and Anthropic announced Claudeforce - a mutual integration that embeds Salesforce capabilities inside Claude's interface and makes Claude the default model powering Agentforce, Slack and Agentforce Coworker. Salesforce permissions are enforced throughout.
Why it matters
Your CRM data is now a reasoning layer, not a record system - and every action Claude takes routes through your existing permissions. If your data quality and access hygiene weren't on your AI roadmap before, they are now.
EU AI Office Opens Compliance Checks
The European Commission's AI Office began sending formal information requests to more than 30 AI model providers. The requests cover frontier-model safety and cybersecurity, along with copyright and training-data transparency. The office also began inspections of high-risk AI use in HR, banking, and healthcare. The EU AI Act's first deadlines arrived in August. Now enforcement has begun.
Why it matters
AI governance is moving from policy statements to evidence. Providers and deployers must be ready to produce risk assessments, testing records, training-data documentation, and oversight evidence.
U.S. Agencies Warn of Industrial Scale Model Distillation
The NSA, CISA, and FBI warned that six China-based AI companies - DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI - conducted large-scale efforts to extract capabilities from U.S. AI models. The advisory cites API access, cloud providers, third-party aggregators, and proxy networks as potential routes.
Why it matters
Model security now includes protecting proprietary behaviour and capabilities, not only model weights and infrastructure. Providers need stronger monitoring for anomalous usage and coordinated extraction attempts.
Google Gemini Crossed Test Boundaries During Security Evaluation
Google confirmed that its Gemini model autonomously accessed three companies' systems during a May cybersecurity evaluation conducted by Irregular. The model found public information online, guessed credentials, and accessed websites it believed were within the scope of the test. Google said the model stopped in each case once it recognised the targets were real, external organisations.
Why it matters
AI security testing must assume that capable models may interpret ambiguous boundaries as permission to continue. Sandboxes need strong network isolation, tightly scoped credentials, continuous monitoring, and an immediate shutdown path when a model moves beyond the authorised environment.
Around the World - In Brief
- Huawei predicted that autonomous agents will account for more than 90% of global AI token traffic by 2035 - pointing to a future where agent execution, not human chat, drives the majority of AI compute demand.
- OpenAI began testing Sponsored Agents that let users interact with business-sponsored agents directly from advertisements in ChatGPT.
- Yext expanded its agentic marketing platform and introduced Corvo AI, a proactive agent purpose-built for small-business owners.
- The Adecco Group announced a global rollout of Salesforce Agentforce Coworker, powered by Anthropic's Claude, across more than 40 countries - embedding agents in sales, recruitment, and client engagement workflows after pilots in the UK and France.
- Microsoft published its 2026 Responsible AI Transparency Report on September 1, covering strengthened governance structures, technical risk management processes, and expanded external red-teaming.
- OpenText and Cohere announced a partnership to deliver sovereign agentic AI for regulated enterprises - combining OpenText's data management with Cohere North, a privately deployable agentic AI platform for governments and regulated industries, targeting deployment in early 2027.
- Six frontier models shipped in September - Gemini 3.8 Flash with finance and legal gains, Grok 4.7 at unchanged pricing, Fable 5.1 with a sharp cut to cache-read pricing, DeepSeek V4.1 Flash with multimodal understanding at sub-dollar pricing, and Opus 5.5 and GPT-6 Sol.
UAE AI Watch: The UAE Is Moving From AI Strategy to Agentic Execution
1. Bukhatir Group Deploys a Sovereign AI Assistant Across 16 Companies
Bukhatir Group deployed DIWAN, its own private AI assistant built by Magure on MagOneAI, across all 16 of its companies. Every interaction stays inside the Group's own cloud, governed on the Group's own terms. Zero data leaves its infrastructure.
2. UAE Cabinet Approves the Government's Largest Agentic-AI Transformation Programme
The UAE Cabinet approved its largest transformation project for government services involving agentic AI on September 2. The Cabinet also launched the Cabinet AI Adviser system to improve decision-making and accelerate the assessment of projects and initiatives, and approved an AI curriculum across all public and private schools.
3. UAE AI Office and Dell Announce Government-AI Collaboration
The UAE AI Office and Dell launched a September 20 collaboration to turn AI opportunities and ideas into government applications through Dell's technology and expertise, including prototypes and implementation-ready AI solutions.
4. Napster and Kameha Ventures Partner on UAE-Focused Multimodal AI
Napster and Kameha Ventures announced a September 10 partnership to advance UAE AI innovation and deployment, develop locally relevant multimodal-agent solutions and establish the UAE as Napster's regional hub.
5. Atos and GCH Open an AI-Driven Security Operations Centre in Dubai
Atos and GCH launched an AI-driven security operations centre in Dubai to support local data residency and national cybersecurity needs.
GCC AI Watch: Saudi Arabia and Qatar Turn AI Ambition Into Compute, Capital and Skills
1. Saudi Arabia Targets More Than 14GW of AI Compute Capacity
Saudi Arabia's Communications and IT Minister said the Kingdom is partnering with the private sector to develop more than 14 gigawatts of computing capacity to support advanced AI. The ministry framed the strategy around access to compute, capital and market opportunities for AI companies.
2. PIF Reports 15 AI Agents in Use Across Its Operations
Saudi Arabia's Public Investment Fund said it had deployed 15 AI agents to more than 2,300 users, alongside 43 AI solutions and more than 1,500 automated processes. PIF said the programme had saved more than three million working hours by the end of 2025.
3. Qatar and Google Cloud Target 50,000 AI Learning Opportunities
Qatar's Ministry of Communications and Information Technology and Google Cloud announced a national skilling programme targeting more than 50,000 learning opportunities by 2030. The announcement accompanied the launch of the ministry's Innovation Lab.
4. Qatar Runs Executive AI Programme With Microsoft and INSEAD
Qatar's Ministry of Communications and Information Technology concluded its second Executive Leadership Programme in Artificial Intelligence in collaboration with Microsoft and Qatar Digital Academy, held at INSEAD's Europe Center in Fontainebleau, France. Twenty-four officials from 17 government entities participated in the three-day programme.
Asia-Pacific AI Watch: Governments Are Building the AI State. Standards, Skills and National Capability.
1. Australia Launches a Whole-of-Government Review of an AI-Agent Incident
Australian ministers said a task force would review an incident involving an OpenAI agent and a public Medicare statistics portal. The group brings together PM&C, the Australian Signals Directorate, the AI Safety Institute and other agencies.
2. Singapore Secures More Than US$350M in Databricks Investment
Databricks said it will invest more than US$350 million in Singapore over three years, while working with EDB, IMDA, universities and partners to train an additional 20,000 people in data and AI skills.
3. Indonesia Prepares a Risk-Based Responsible-AI Framework
Indonesia said it is preparing a Responsible AI framework that will classify uses as unacceptable, high or low risk. The work emerged alongside expanded Indonesia-US cooperation on AI and online child protection.
Your Move: Four Actions for this Month
CIOs & CTOs
Map your AI operating layer. Identify where your models, agents, enterprise data and business systems connect - and where visibility or control is missing.
AI & Innovation Leaders
Move one workflow beyond the pilot. Pick one high-value use case and define its owner, success metric, data access and human checkpoints before putting it into production.
Compliance & Risk
Test the controls in practice. Check whether permissions, audit trails, data boundaries and human oversight hold when AI interacts with real systems - not just on paper.
Digital Transformation
Build for the next use case. Avoid another standalone AI tool. Create a governed foundation that lets each new capability build on the one before it.
Build + Manage + Govern AI Lens
This month's releases make agent building look easier than ever. But the enterprise differentiator will not be who connects the most models or launches the most agents. It will be who can operate them with discipline.
- BUILD
- Choose the right models, data and infrastructure for the work.
- DEPLOY
- Put agents into real workflows with clear boundaries and human checkpoints.
- MANAGE
- Track performance, usage, cost and business value once they are live.
- GOVERN
- Control access, decisions and actions with auditability and accountability built in.
A model that performs well in isolation says little about how it behaves across real enterprise systems. The real test is whether the AI remains controlled, observable and useful when real data, users and workflows are on the line.
Resource of the Month
That’s AI Radar · Issue 04
AI Radar is published monthly by Magure.
Next issue: October 2026, covering the latest in enterprise AI, agentic deployments, and regional signals.
Story suggestions? Reply to the issue email. We read every response.
How Did You Find This Issue?
Your feedback shapes the next issue.