AI Leadership Weekly · Issue #81 · Tuesday 28 April 2026 · 08:00 BST
Good morning.
OpenAI introduced GPT-5.5. Google announced its enterprise agent platform, and Anthropic added memory to Managed Agents. Memory is the particularly useful thread for business readers: information can persist beyond a single interaction, but someone still needs to decide what it means and when it becomes stale. The platform announcements deserve attention alongside the model headlines, not underneath them.
IN 60 SECONDS
GPT-5.5 launches. OpenAI introduced GPT-5.5, broadening its model offering for professional and coding work; API availability followed the initial announcement.
Google introduces an agent platform. Google announced Gemini Enterprise Agent Platform, bringing development, deployment and management of enterprise agents into a connected offering.
Managed Agents gains memory. Anthropic added inspectable, persistent memory for Claude Managed Agents, letting developers manage information carried between sessions.
CEO / COO / CXO CHECKLIST
CEO: Distinguish authoritative policy from observations saved during previous work.
COO: Decide who can inspect, correct and remove persistent information.
CXO: Include an outdated instruction in your agent tests and check what takes precedence.
TOP STORIES
1. GPT-5.5 launches
OpenAI · 23 April 2026
What happened. OpenAI introduced GPT-5.5 on 23 April for work including coding, research, document production and computer use. ChatGPT and Codex rollout began with the announcement; API availability followed on 24 April. The release reports performance improvements, but those are not a guarantee of better outcomes in every existing process.
Why it matters. Our take: A broader capability set makes it tempting to combine several steps into one instruction. That can remove handovers, but it can also conceal intermediate mistakes. Decide which checkpoints remain necessary. A long task should still have an understandable result, a bounded remit and a way to show what happened when it diverged from the request. Judge the whole completed job, including review, rather than its most impressive stage.
What to do. Select a task that currently crosses two tools. Trial a combined approach while retaining the existing checks. Record whether it removes work or simply moves that work into a longer final review.
2. Google introduces an agent platform
Google · 22 April 2026
What happened. Google announced Gemini Enterprise Agent Platform on 22 April, bringing together model services and capabilities for building, running and governing agents. It described support for its own and partner models. The platform announcement does not mean every component or configuration was generally available to every customer at launch.
Why it matters. Our take: An integrated platform may reduce the effort of assembling separate components. It also makes the buying decision broader than model quality. Ask how identity, operational records, tests and data connections travel together. Then ask what can be exported or replaced. The goal is to obtain useful integration without losing visibility into the parts on which the service depends.
What to do. Ask a prospective supplier to walk through one failed task, not only a successful one. Follow its permissions, evidence, cost and recovery route. Note which records belong to your organisation and how you would retrieve them.
3. Managed Agents gains memory
Anthropic · 23 April 2026
What happened. Anthropic announced memory for Claude Managed Agents on 23 April in public beta. The service allows persistent information to be inspected and managed, with controls including scoped access and a history of changes. The availability of these controls does not establish that a saved conclusion is correct.
Why it matters. Our take: A useful memory policy needs to distinguish a reusable lesson from an accidental preference or one-off concession. For example, remembering the preferred format of a report is different from remembering that a particular exception was once approved. Give authoritative policy priority and make conflict visible. Otherwise an assistant can confidently repeat an old decision after the relevant rule has changed.
What to do. Review a proposed memory store with a process owner. Mark what may be retained, what needs a time limit and what must never become a general instruction. Test correction as carefully as saving.
SIGNALS FROM THE LAST MONTH
8 April · Claude Managed Agents enters beta. Anthropic introduced a public beta for managed agent infrastructure, taking on more of the execution and session-management layer. Source
8 April · Colab adds a learning mode. Google introduced Colab features aimed at helping people understand and develop code, rather than only generating finished answers. Source
7 April · Project Glasswing focuses on defence. Anthropic announced Project Glasswing with selected partners, using restricted model access to find and address software vulnerabilities. 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
IN BRIEF
More dated updates from the preceding 30 days.
21 April · Google updates Deep Research. Google announced its next generation of Gemini Deep Research, extending its developer offering for multi-step research tasks. Source
20 April · Google AI subscriptions connect to AI Studio. Google introduced a route for eligible AI subscribers to prototype in AI Studio, subject to the service’s usage conditions. Source
16 April · Claude Opus 4.7 is released. Anthropic launched Opus 4.7, with changes in capability and model behaviour that make migration testing important. Source
15 April · IBM responds to agentic cyber threats. IBM announced coordinated security assessment and automation measures aimed at helping enterprises respond to AI-enabled attacks. Source
THE 15-MINUTE PLAYBOOK
Minutes 0–4: List five things the assistant might remember. Include a preference, a factual record, a policy, a previous decision and an unresolved assumption.
Minutes 4–8: Assign an owner and an appropriate lifetime to each. Identify which must be refreshed from an authoritative source rather than trusted indefinitely.
Minutes 8–12: Create a conflict: an old saved instruction and a new approved rule. Specify which wins and what the user should be told.
Minutes 12–15: Define how a colleague requests correction, how related records are found and who verifies that the correction changed subsequent behaviour. Keep the first policy narrow enough to test on real work.
DATA WAVE MOMENT
Organisational knowledge becomes useful when it is connected to a clear task and can be maintained by the people responsible for it. Start with a small body of approved information, a visible correction process and a specific outcome. Let the memory grow through deliberate use, not accidental accumulation.
QUESTION FOR READERS
What would an assistant keep doing incorrectly if it remembered last month perfectly?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
