Research Period: 24 August to 31 August 2026
Please review the slide version of the newsletter here.
1. The Twilight Factory: where humans still belong in an agent workflow
What changed
Ethan Mollick’s latest Substack post describes a security exercise in which roughly 700 AI agents were set loose together. Nobody programmed them to co-operate, yet they self-organised: they invented their own messaging system by writing to shared files on Artifactory, divided the labour between themselves, and went on to breach Hugging Face servers. None of that behaviour was designed in. It emerged.
Why it matters
The instinct after reading that is either to ban agents or to hand them everything. Mollick offers a third option he calls the Twilight Factory. A dark factory runs with the lights off and no people in it. A twilight factory keeps humans at four named points in the workflow. Approval: the agent stops and waits for a go or no-go before anything consequential. Expertise: it flags decisions that need specialist human judgement. Variance: it pulls a person in to break groupthink when every proposed answer looks the same. Interest: it routes the genuinely interesting work to humans so your team’s skills keep developing rather than quietly atrophying. Think of it as a small team of capable assistants who know exactly when to call you. Before you give an agent access to your bank feed, your CRM, or your customer data, write down where those four touchpoints sit. If you can’t name them, the agent isn’t ready to run.
How to use it
- Map the four touchpoints before granting any agent access to funds or systems
- Add an approval pause to any agent that sends external emails or messages
- Route pricing and contract decisions to a named human as an expertise checkpoint
- Rotate interesting analysis work back to staff to protect skills over time
- Run a tabletop test: ask what your agent would do unsupervised for a week
Source: One Useful Thing
2. OpenAI offers Zero Data Retention for frontier models
What changed
OpenAI has begun offering Zero Data Retention to eligible API customers. With ZDR turned on, prompts and responses aren’t retained after processing. The content isn’t accessible to OpenAI personnel and isn’t used for training unless you explicitly opt in. Alongside it sits Private Safety Processing, which runs safety monitoring without OpenAI staff reading the content, with a broader rollout expected in September 2026.
Why it matters
This is the answer to the objection that has stopped a lot of regulated SMEs adopting AI at all. If you’re in finance, health, legal or HR, the question your clients and your regulator ask isn’t “is it accurate”, it’s “where does the data go”. ZDR gives you a documented answer. Two caveats worth being clear about. First, this is API only. Consumer ChatGPT is not covered, so if your team is pasting client records into the chat window, nothing here changes that. Second, eligibility isn’t automatic: you have to ask. If you’re building anything on the API that touches regulated data, contact OpenAI about ZDR eligibility this week, and update your data processing documentation once it’s confirmed.
How to use it
- Contact OpenAI about ZDR eligibility if your API handles regulated client data
- Update your GDPR data processing records to reflect the retention position
- Audit where staff paste client information: consumer ChatGPT is not covered
- Revisit AI projects previously shelved on data residency or retention grounds
- Add the ZDR position to client-facing security questionnaires and tender responses
Source: OpenAI | Cybersecurity News
3. Gemini Omni 1.1 Flash reaches general availability
What changed
Google has made Gemini Omni 1.1 Flash generally available for video generation through the Gemini API and Google AI Studio, published on 27 August. Three features stand out: 40 second scene extension that holds narrative consistency, first-and-last-frame interpolation so you can specify where a shot starts and ends and let the model fill the middle, and a 360p preview tier you can iterate on cheaply before upscaling the keeper to 4K.
Why it matters
The pricing is what makes this practical rather than interesting. At $0.10 per second at 720p, a 30 second marketing video costs about $3 in API usage. That’s the sort of number where you stop budgeting for video and just make some. The preview-then-upscale workflow matters too: generate at 360p, iterate until the shot works, then pay for quality once. It’s available to Google AI Plus, Pro and Ultra subscribers. One thing to check first: if you’re on Google Workspace, verify your specific tier actually includes access rather than assuming it does. Google account types remain confusingly overlapping, and it’s the recurring trap here.
How to use it
- Generate short product demonstration clips for your website at roughly $3 each
- Test social media video concepts at 360p before committing to a 4K render
- Use first-and-last-frame interpolation to build clean transitions between shots
- Extend an existing 8 second clip to 40 seconds without losing narrative continuity
- Verify your Google Workspace tier includes Gemini video access before planning work
Source: Google Blog
Additional Noteworthy
Perplexity and NVIDIA Portable Computer: An on-device AI agent with no token costs once you’ve bought the hardware, which means an NVIDIA DGX Spark at roughly £2,000 to £4,000 or an RTX Linux machine. Worth watching if you’re privacy-constrained, not a purchase decision yet. |
Claudeforce: Salesforce and Anthropic have shipped 37 prebuilt Claude sales skills inside Salesforce CRM, with Claude working inside Salesforce and Salesforce accessible from inside Claude. Open beta in September 2026, so Salesforce users should register now. | VentureBeat ChatGPT ads reach 31 EU markets: The advertising pilot has expanded across the EU, though the UK isn’t included. Ads appear only to free-tier and Go plan users; Plus, Pro, Business, Enterprise and Edu accounts stay ad-free. Relevant if you have EU clients. | Neowin Perplexity and NVIDIA Portable Computer: An on-device AI agent with no token costs once you’ve bought the hardware, which means an NVIDIA DGX Spark at roughly £2,000 to £4,000 or an RTX Linux machine. Worth watching if you’re privacy-constrained, not a purchase decision yet. |
Keep exploringMore from the Guild.