AI Leadership Weekly · Issue #79 · Tuesday 14 April 2026 · 08:00 BST
Good morning.
Anthropic’s two announcements this week concern very different kinds of access: a public beta for managed agent infrastructure and a restricted programme for defensive software-security work. Google’s Colab update puts learning alongside code generation. Together they make a useful counterpoint to the assumption that every release is simply a more powerful model that everybody can immediately use.
IN 60 SECONDS
Claude Managed Agents enters beta. Anthropic introduced a public beta for managed agent infrastructure, taking on more of the execution and session-management layer.
Project Glasswing focuses on defence. Anthropic announced Project Glasswing with selected partners, using restricted model access to find and address software vulnerabilities.
Colab adds a learning mode. Google introduced Colab features aimed at helping people understand and develop code, rather than only generating finished answers.
CEO / COO / CXO CHECKLIST
CEO: Write down what the reviewer must understand, not just what they must approve.
COO: Give new automated work a named route for investigation and recovery.
CXO: Reserve time for learning from exceptions rather than only clearing them.
TOP STORIES
1. Claude Managed Agents enters beta
Anthropic · 8 April 2026
What happened. Anthropic introduced Claude Managed Agents in public beta on 8 April. The service provides building blocks for running agents, including managed execution and the supporting infrastructure for stateful work. It is a way to build an agent application, not an off-the-shelf guarantee that the application will make sound business decisions.
Why it matters. Our take: Buying infrastructure can be sensible when maintaining it is not your differentiator. But make the retained responsibilities explicit: approved sources, task boundaries, tests, access, exceptions and service ownership. A good supplier conversation should explain the boundary between its platform and your application. Without that boundary, every failure risks becoming a discussion about whose layer caused it while the customer waits.
What to do. Draw a two-column responsibility sheet for one proposed agent. Put supplier-operated infrastructure on one side and your business obligations on the other. Resolve any item both parties assume the other owns before expanding the trial.
2. Project Glasswing focuses on defence
Anthropic · 7 April 2026
What happened. Anthropic announced Project Glasswing on 7 April, bringing selected technology and security organisations into defensive work with the unreleased Claude Mythos Preview. The company described vulnerability findings and a programme to help secure important software. This was restricted access, not a generally available model release or independent assurance of security.
Why it matters. Our take: For most leaders, the useful question is not how to obtain the model. It is whether the organisation can act on credible findings. A longer list of weaknesses is only useful when someone can prioritise, test a repair, deploy it and confirm the outcome. Ask about that complete pathway rather than the headline number of discoveries.
What to do. Trace one recent security finding from notification to closure with the responsible team. Note where it waited for a decision or ownership. Use the exercise to improve defensive response, without commissioning unnecessary new scanning.
3. Colab adds a learning mode
Google · 8 April 2026
What happened. Google’s 8 April Colab update introduced Learn Mode, designed to guide users through coding rather than simply produce the answer. It also added notebook-level custom instructions that can travel with the shared notebook. The announcement describes features, not proof of learning outcomes for a workforce.
Why it matters. Our take: Training and production are different contexts. In production, the shortest acceptable route may be valuable. In learning, doing the reasoning is often the point. A useful training exercise should reveal whether the colleague can apply the idea to a different example. Otherwise the organisation may confuse a completed worksheet with a new capability.
What to do. Choose a simple analytical task. Ask a learner to use guided help, then explain the result and adapt it to a changed input. Assess the explanation and the change, not only whether the original notebook ran successfully.
SIGNALS FROM THE LAST MONTH
24 March · ChatGPT improves product comparison. OpenAI announced shopping updates in ChatGPT, including more visual product discovery and side-by-side comparison. Source
24 March · Claude Code previews auto mode. Anthropic introduced a research preview that automates some permission decisions in Claude Code while retaining checks for higher-risk actions. Source
23 March · Mistral expands its voice offer. Mistral announced a Voxtral update, extending its speech-focused model offering beyond text-only interactions. Source
17 March · Personal Intelligence reaches more users. Google expanded Personal Intelligence for eligible US personal accounts. The announcement did not grant equivalent access to Workspace business accounts. Source
IN BRIEF
More dated updates from the preceding 30 days.
2 April · Google Vids adds creation tools. Google expanded Vids with new generative video and music capabilities; feature limits and access depended on the account and plan. Source
2 April · Google releases Gemma 4. Google introduced Gemma 4, adding open-model deployment options for organisations assessing on-device and self-managed AI. Source
31 March · Mistral designs tools for agents. Mistral described Spaces, a command-line tool designed to serve both people and agents working with development environments. Source
26 March · Search Live expands internationally. Google expanded Search Live, allowing more users to ask questions through voice and camera input while searching. Source
THE 15-MINUTE PLAYBOOK
Minutes 0–4: Select one AI-produced output that a colleague is expected to review. Identify the two most consequential ways it could be wrong.
Minutes 4–8: Ask the colleague to explain its sources and assumptions in ordinary language. Notice where the explanation depends on trusting the tool rather than understanding the evidence.
Minutes 8–12: Change one input or constraint. Ask what should change in the result and what should remain stable. Use the difference to identify a training need.
Minutes 12–15: Agree a short review guide and a named escalation route. Repeat the exercise after a few real cases, using actual mistakes to improve the guide rather than adding generic training slides.
DATA WAVE MOMENT
Building from within should leave a team more capable, not merely better supplied with software. Put business judgement, delivery and learning in the same working group. The result should be a useful service and people who can explain its limits, improve its instructions and know when not to use it.
QUESTION FOR READERS
What do our reviewers need to understand that the current approval process never asks them to demonstrate?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
