October 15

Your Daily Note,

AI agents are getting more room to work, but the bigger question is how to keep them useful, predictable and accountable.

Figma is introducing ways to give agents clearer design rules. Ricursive Intelligence is working on AI that can design better chips. Meanwhile, music platforms are figuring out how to handle AI-generated songs, and cybersecurity teams are using AI to investigate threats that once took human experts weeks.

In today’s Noted:

Figma wants AI agents to follow your design rules, not invent their own

We noted:

Figma has introduced Figma Specs, a Markdown-based editor that lets teams document their design language in a format both people and AI agents can reference.

The broader idea is to give Figma’s own AI features and third-party agents a clearer understanding of how a product should look and behave before they start creating.

Instead of repeatedly telling an AI agent to use the right colours, typography, spacing and components, teams could give it a shared reference that explains those decisions. That should make AI-generated designs easier to align with an existing product, rather than producing polished but inconsistent mockups.

Why it matters

Design tools are beginning to define not only what people can create, but also the rules AI agents should follow when creating it.

If you use AI to build landing pages or product interfaces, start documenting your design rules in a reusable file. A good design brief can be more valuable than another round of prompts.

An AI startup thinks the answer to the compute crisis is better-designed chips

We noted:

Ricursive Intelligence is developing AI systems that can design computer chips, learn from each design iteration and use those lessons to improve subsequent designs.

The startup’s founders, Anna Goldie and Azalia Mirhoseini, argue that rising demand for AI computing cannot be addressed simply by building more data centres. More efficient chips and better use of the bandwidth, manufacturing capacity and hardware already available could also help.

Why it matters

The immediate goal is not a chip that appears overnight. Chip design and manufacturing remain complex, and a promising design still needs to be tested and produced. But if AI can shorten design cycles or find improvements that engineers might otherwise miss, chipmakers could have more opportunities to improve performance and efficiency.

Better hardware efficiency can affect how fast and affordably AI tools become available.

Apate sends AI agents undercover to learn how scammers operate

We noted:

Australian startup Apate uses AI agents that pretend to be potential victims, engaging with suspected scammers to collect intelligence about their operations. Founded in 2024 from research at Macquarie University, the company deploys agents that can respond to suspicious calls, messages and emails.

According to the company, its agents have accumulated 2.5 million conversations and wasted a combined 17 years of scammers’ time.

Why it matters

Banks, telecommunications companies, government agencies and utilities can use the resulting intelligence to understand scam patterns and identify connections between apparently separate incidents. Rather than relying only on reports from people who have already encountered fraud, organisations can gain additional information from proactive investigations.

For businesses, fraud prevention is not only about detecting suspicious messages. Tracking impersonation, recurring websites and repeated contact details can help reveal a wider campaign.

YouTube Music is adding labels to AI-generated songs

Image credit: YouTube Music

We noted:

YouTube Music is rolling out labels that help listeners identify music made entirely or partly with generative AI. The labels will also extend to the main YouTube app.

The labels appear on a track’s player page, where listeners can select them to learn more about AI’s role in its creation. YouTube is working with music partners and using its own detection tools to identify relevant tracks.

Why it matters

Listeners should have more context about how a song was made. Artists and music partners will also need to pay closer attention to how they disclose AI use when uploading tracks.

The labels are not a guarantee that every AI contribution will be identified perfectly, but they provide a clearer starting point for transparency. Creators should keep track of how their tools contributed to the final work, since disclosure requirements can differ between platforms.

OpenAI says SentinelOne cut malware analysis from weeks to hours

We noted:

In a newly published customer case study, OpenAI says cybersecurity company SentinelOne used its Daybreak capabilities to reduce the time required for certain malware reverse-engineering tasks from four to six weeks to roughly eight hours.

Reverse engineering involves examining software to understand what it does and how it behaves. For security researchers, that can mean painstakingly investigating suspicious programs, tracing their behaviour and identifying links to other threats.

Why it matters

Faster analysis could help security teams examine more threats instead of reserving intensive investigation for only the most urgent cases.

The eight-hour figure is a result reported in OpenAI’s customer case study, not a guarantee that every malware investigation can be completed in that time. If a business uses AI for security, evaluate whether it improves the full investigation process, including verification and follow-up, rather than measuring success by response speed alone.

Products worth trying

 

Everything else in AI

Deezer is preparing a filter for AI-generated music

Music-streaming service Deezer says it is developing a setting that will let listeners remove AI-generated tracks, albums and related artists from search results. The feature is expected before the end of 2026.

The company reported that more than half of its daily music uploads were AI-generated in July, averaging around 90,000 tracks per day. That figure reflects Deezer’s own reporting, but it illustrates the scale of the content-management challenge facing streaming services.

YouTube is leaning towards disclosure, while Deezer is moving towards giving listeners an explicit choice to filter this content out.

 

Gudea wants to spot online reputation crises before they erupt

Gudea analyses how stories and rumours spread across social platforms and websites. Rather than simply counting mentions or hashtags, its system aims to identify who is driving a conversation, how information travels and whether a narrative is gaining momentum.

The company says it monitors more than 500 mainstream and fringe websites and platforms, including X, Reddit and TikTok. It has nearly 40 clients, according to its CEO, and recently announced a $7 million seed round.

For organisations, the promise is earlier warning and more context before responding to a fast-moving controversy. The challenge will be distinguishing genuinely emerging risks from ordinary online discussion without overreacting to every spike in attention.

 

Flock’s microphone claims raise questions about surveillance and transparency

Flock CEO Garrett Langley said the company’s audio-detection product was designed to detect gunshots, but TechCrunch noted that the company’s website describes detection capabilities that extend to other events, including fireworks, vehicle takeovers and potential car crashes.

The company says the technology is not designed to monitor conversations or analyse speech. The gap between the CEO’s explanation and the broader descriptions of the product has renewed questions about how audio surveillance systems are described, what they can detect and how their capabilities might expand.

The wider lesson for AI-enabled monitoring is that clear limits, transparent documentation and meaningful oversight matter just as much as the technology’s intended use.

 

Zenzap raises $44 million for AI-powered workplace messaging

Zenzap is building a workplace messaging app intended to replace consumer chat tools that businesses often use to coordinate staff and exchange documents. Its pitch is to combine messaging with AI that understands organisational context while keeping communication within company security and IT rules.

The company raised $44 million from investors including Bessemer and Lightspeed. The opportunity is not simply adding a chatbot to workplace chat. It is making company knowledge and everyday coordination more accessible without forcing employees to search across disconnected conversations and documents.

Whether businesses switch will depend on integrations, adoption and how well the product handles sensitive company information.

 

Floqer wants AI agents to become a continuously updated customer database

Floqer is building an autonomous customer-knowledge platform for business-to-business sales and marketing teams. Its agents collect, organise and update information that revenue teams use to understand prospects and customers.

The startup says it has passed $1.2 million in annual recurring revenue and plans to open-source the technology behind its customer-knowledge database in early 2027, pending security reviews.

The idea is to reduce the manual research required before a sales conversation or campaign. The harder question is whether the information remains accurate, well sourced and current enough to support decisions.

 

Third Moment Research is running thousands of AI agents on long research tasks

In a separate OpenAI customer case study, Third Moment Research describes a system that runs up to 10,000 agents concurrently on research sessions lasting 36 to 48 hours. The company uses OpenAI models and Codex to investigate problems, run experiments and improve its agent workflows.

One example involved optimising a financial backtesting process that had taken roughly three hours. The company says a change identified with Codex reduced that process to around three to five minutes.

These are company-reported results from a specific research environment, not a general guarantee of performance. Still, the case illustrates how AI agents can be organised into longer-running workflows that involve experimentation, debugging and follow-up work rather than a single prompt and response.

Worth your weekend

See you tomorrow
same time, different note.

~ the Stackari team

 

P.S. One thing at a time, one note at a time. Just one more day till Friday!