OpenAI rolls out Presence, its enterprise agent platform.

In partnership with

Write docs 4x faster. Without hating every second.

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

Yep, the big three are at it again!

Forward thinkers, welcome to issue #168 of the Neural Frontier! 

Today, we’re unpacking major moves from OpenAI, Google, and Anthropic. Ready or not, here we go! 

In a rush? Here's your quick byte: 

🧑‍💻 OpenAI rolls out Presence, its enterprise agent platform.

🚀 Google launches three new Gemini models.

👀 Claude can now learn your workflows by watching you work!

🎭 AI Reimagines: The Lord of the…

🎯 Everything else you missed this week.

⚡ The Neural Frontier’s weekly spotlight: 3 AI tools making the rounds this week.

🧑‍💻 OpenAI rolls out Presence, its enterprise agent platform.

OpenAI wants to move AI agents out of the demo stage and into the messier world of real business operations.

The company has introduced OpenAI Presence, an enterprise product designed to help companies deploy voice and chat agents across customer-facing and internal workflows.

These agents can answer questions, access company systems, take approved actions, and hand conversations over to human employees when necessary.

🤖 What OpenAI Presence actually does

Presence combines OpenAI’s models with the infrastructure businesses need to manage agent behavior.

Instead of giving an agent broad access to company systems, each deployment begins with a specific job, such as:

• Resolving customer billing issues.

• Supporting insurance claims.

• Handling employee IT requests.

• Managing outbound sales conversations.

The agent receives only the information, permissions, and system access required for that particular task. Companies can then define exactly how it should operate, including what actions it can take independently, when approval is required, and when a human employee needs to step in.

🛡️ Built around control, not complete autonomy

OpenAI is positioning Presence as a way for businesses to use agents without giving up control over sensitive workflows.

The platform includes:

• Company policies and standard operating procedures.

• Guardrails and approved actions.

• Simulations and evaluation tools.

• Human approval and escalation rules.

• A Codex-powered process for improving agents after launch.

After an agent goes live, companies can review production conversations and human escalations to identify where it is struggling. Codex can then propose changes to the agent’s instructions, policies, or workflow. Teams can test those changes and approve them before they are deployed.

This creates an improvement loop that helps the agent adapt as products, customer behavior, and company policies change.

☎️ Available for voice and chat agents

Presence currently supports real-time voice and chat experiences, particularly in areas like customer service, outbound sales, and sensitive internal operations.

Businesses can reuse the same policies, evaluation systems, and escalation rules across different channels while customizing the agent for each workflow. That could make it easier for companies to expand from one successful deployment into several related use cases without rebuilding everything from scratch.

🚀 Google launches three new Gemini models.

Google is expanding the Gemini lineup with three new models designed to make AI agents faster, cheaper, and more reliable in production.

The company has launched Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, each targeting a different part of the agentic AI market.

🧠 Gemini 3.6 Flash is the new workhorse

Gemini 3.6 Flash is positioned as Google’s main model for developers building production AI agents. It improves on Gemini 3.5 Flash across coding, knowledge work, computer use, and multimodal tasks, while using fewer tokens to complete the same work.

According to Google, the model uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index. In some coding benchmarks, the reduction is significantly larger.

Google says the model performs particularly well in:

• Coding and software migration tasks.

• Document parsing and report writing.

• Chart, image, and financial-data analysis.

• Computer-use workflows involving websites and applications.

On DeepSWE, a software-engineering benchmark, 3.6 Flash scored 49%, compared with 37% for 3.5 Flash. It also improved on OSWorld-Verified, which measures an AI system’s ability to operate computers, scoring 83% versus 78.4%.

🏎️ Gemini 3.5 Flash-Lite prioritizes speed and scale

For developers who care more about volume and latency than maximum intelligence, Google has introduced Gemini 3.5 Flash-Lite. It is the fastest model in the Gemini 3.5 family, generating up to 350 output tokens per second according to Artificial Analysis.

The model costs $0.30 per million input tokens and $2.50 per million output tokens, making it significantly cheaper than 3.6 Flash.

Google is positioning Flash-Lite for tasks such as:

• Agent-powered search.

• Large-scale document processing.

• Translation and summarisation.

• Data extraction and classification.

• Running multiple smaller agents beneath a more capable primary model.

Developers can also adjust how much reasoning the model uses. Lower thinking levels are intended for fast, repetitive tasks, while higher settings can be used for more complex multi-step workflows.

🛡️ Gemini 3.5 Flash Cyber focuses on software security

The third release is more specialized. Gemini 3.5 Flash Cyber is a version of 3.5 Flash fine-tuned to identify, validate, and fix cybersecurity vulnerabilities.

The model is being deployed through CodeMender, Google’s code-security agent, where several Flash Cyber agents work together to analyze software and produce a combined report.

Google says this approach delivers competitive performance on CyberGym, a benchmark for evaluating an AI system’s ability to find and patch security flaws.

But access will be tightly restricted. Because cybersecurity models can be used for both defensive and offensive purposes, Flash Cyber will initially be available only to governments and selected trusted partners through a limited pilot.

👀 Claude can now learn your workflows by watching you work!

Explaining a complicated workflow to an AI can sometimes take longer than simply doing it yourself. Anthropic wants to fix that.

Claude has introduced a new “Record a skill” feature that lets users teach the chatbot how to complete a task by recording their screen and talking through the process.

Instead of writing a lengthy prompt describing every step, users can demonstrate the workflow directly. Claude watches where they click, what they type, and how they move through different tools, then turns the recording into a reusable skill it can run again.

🖥️ Show Claude instead of explaining everything

The feature is part of Claude Cowork, Anthropic’s suite of tools for completing more complex work across files and applications.

To create a skill, users can:

• Start a short screen recording inside the Claude desktop app.

• Complete the task as they normally would.

• Explain important decisions through a voiceover.

• Let Claude convert the demonstration into a repeatable workflow.

Claude uses the recording to understand both the actions being taken and the reasoning behind them. Once Claude processes the recording, it creates a skill that can be reused whenever the same task needs to be completed.

📊 Built for repetitive workplace tasks

Anthropic is positioning the feature for processes that employees regularly repeat but struggle to document clearly.

For example, someone who prepares a weekly report could show Claude:

• Where to retrieve information from different platforms.

• Which data points should be included.

• How the information should be organised.

• What the final report should look like.

Claude could then reproduce that process without requiring the user to explain every step again.

The feature could also make it easier for companies to preserve operational knowledge. Instead of relying entirely on written standard operating procedures, employees could demonstrate how a process works and allow Claude to turn it into something reusable.

⚡ The Neural Frontier’s weekly spotlight: 3 AI tools making the rounds this week.

1. 🔍 AnySearch is an AI search infrastructure that gives agents access to high-quality, structured data across multiple domains through a single API..

2. 👀 isvisible.ai is a free AI visibility audit tool that checks whether 13 major AI and search crawlers can access your website, then provides a 0–100 score, crawler-by-crawler breakdown, and downloadable report.

3. 🌐 Context.dev is a web data API that gives AI agents live, structured context from any website, handling scraping, crawling, screenshots, sitemaps, and brand or product extraction through one integration.

Wrapping up… 

The thread across this week’s updates is pretty clear: agents are getting closer to the actual workflows companies care about.

None of this is as flashy as a brand-new frontier model, but it’s just as important. The real test for agents won’t be whether they look impressive in demos, but whether companies can trust them inside everyday work.

See you next week on The Neural Frontier 👋

PS: If you’re the “friend” this mail was forwarded to, and you enjoyed it, hit the Subscribe button to see more content like this every week 🙂.