AI Leadership Weekly · Issue #85 · Tuesday 26 May 2026 · 08:00 BST
Good morning.
Google released Gemini 3.5 Flash, while Anthropic and Mistral expanded how agent work can run. The product names are changing quickly, but the underlying decision is familiar: which work belongs where? A remote agent, a self-hosted execution environment and a managed platform place different responsibilities on your team. The useful comparison is the finished task, not the number of agents involved.
IN 60 SECONDS
Gemini 3.5 Flash launches. Google introduced Gemini 3.5 Flash while describing Pro as forthcoming, making the distinction between released and planned models important.
Managed Agents adds deployment choices. Anthropic introduced self-hosted sandboxes and MCP tunnels, allowing organisations to separate aspects of execution from managed orchestration.
Mistral introduces remote agents. Mistral added remote agents to Vibe and released Medium 3.5, widening its offer for delegated development work.
CEO / COO / CXO CHECKLIST
CEO: Replace a vague instruction with a specific deliverable and acceptance checks.
COO: Separate read access, write access and permission to communicate externally.
CXO: Set a stopping condition for missing evidence, rising cost or an ambiguous decision.
TOP STORIES
1. Gemini 3.5 Flash launches
Google · 19 May 2026
What happened. Google introduced the Gemini 3.5 family on 19 May, launching Flash across its products and developer platforms. The announcement described Pro as still internal and planned for later release. Treat Flash’s launch separately from the capabilities and availability promised for the wider family.
Why it matters. Our take: Product-family announcements can blur the difference between what a team can trial now and what might become available later. Keep delivery decisions anchored to the actual model, account and deployment route being tested. Then assess the task as a whole: did the system use appropriate sources, finish within its remit and produce something a reviewer could accept without reconstructing the work?
What to do. Add exact model and availability details to one trial plan. State the work it is allowed to perform and the evidence required to accept the result. Leave future releases in a separate options column.
2. Managed Agents adds deployment choices
Anthropic · 19 May 2026
What happened. Anthropic announced self-hosted sandboxes in public beta and MCP tunnels in research preview on 19 May. The sandbox option lets tool execution take place in a customer-controlled environment, while the agent loop remains on Anthropic infrastructure. It is not an announcement that the entire service runs privately on premises.
Why it matters. Our take: “Runs in our environment” needs a more precise diagram. Which data travels to the model? Where are tool outputs stored? Which credentials can be used? Who can inspect the execution? These questions determine whether the design fits your information rules. A useful architecture discussion should show the actual path of one request, not rely on a single reassuring label.
What to do. Draw that path for a proposed task, including prompts, source extracts, tool outputs and logs. Ask the relevant owners to confirm each boundary before approving sensitive information for the trial.
3. Mistral introduces remote agents
Mistral AI · 22 May 2026
What happened. Mistral announced Medium 3.5 and remote Vibe agents on 22 May. The model was released in public preview with open weights under modified MIT terms. Remote coding sessions can run in cloud sandboxes and return work for review; the announcement also introduced a preview of broader Work mode.
Why it matters. Our take: The useful shift is from supervising every keystroke to reviewing a completed unit of work. That only works when the unit is well defined. A dependency update with clear tests is different from “improve the application”. For non-coding work, the equivalent is a specified report, reconciled dataset or prepared set of options—not an unrestricted instruction to resolve a business problem.
What to do. Choose a bounded task with a reversible output. Specify its tests and ask for a summary of changes and unresolved issues alongside the deliverable. Keep final acceptance with the appropriate owner.
SIGNALS FROM THE LAST MONTH
7 May · AlphaEvolve reports further applications. Google described additional uses of AlphaEvolve for algorithmic optimisation, extending AI applications beyond conversational assistance. Source
6 May · Uber describes its AI assistant. An OpenAI customer account described Uber’s use of AI to connect assistance with the information and decisions involved in operating its services. Source
5 May · Microsoft studies new working patterns. Microsoft described different patterns of collaboration between employees and agents, focusing on the organisation of work rather than a single product launch. Source
4 May · IBM surveys changing leadership roles. IBM’s study of 2,000 CEOs and equivalent leaders reported changes in AI responsibilities; its results describe that surveyed population. Source
IN BRIEF
More dated updates from the preceding 30 days.
19 May · Google develops a universal shopping cart. Google announced a universal shopping-cart experience, extending its work on AI-assisted discovery and purchasing. Source
15 May · Databricks tests enterprise agent workloads. Databricks described GPT-5.5 results on enterprise document tasks. The published comparisons were supplier-reported evaluations, not customer-wide guarantees. Source
14 May · IBM proposes integrated AI delivery teams. IBM described small, integrated delivery units intended to connect domain knowledge, engineering and implementation around specific business problems. Source
12 May · IBM adds managed Red Hat inference. IBM announced Red Hat AI Inference and OpenShift Virtualization services on IBM Cloud, expanding enterprise deployment options. Source
THE 15-MINUTE PLAYBOOK
Minutes 0–4: Choose one task you would like to delegate. Describe its output and the person who will use it. Avoid a broad objective such as “improve efficiency”.
Minutes 4–8: List approved sources and permitted actions. State explicitly whether the agent may edit shared records or contact another person.
Minutes 8–12: Define acceptance checks, a time or spend limit and conditions that require clarification. Include what to do when two sources disagree.
Minutes 12–15: Write the handover format: result, evidence, changes, checks and open questions. Trial it on a small example before granting recurring or wider authority.
DATA WAVE MOMENT
Effective delegation begins with understanding the work well enough to set its boundaries. Build the first task with the business owner, make the result easy to inspect and preserve a simple route back to a person. Expand only when the evidence shows that the next level of responsibility is justified.
QUESTION FOR READERS
Which of our current AI instructions would be too vague to give to a new colleague?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
