AI agents have quietly accumulated detailed behavioral records of users, which can now be extracted and analyzed using a publicly available prompt called the Reflection Engine.
When uploaded to agents like Sol Max, the tool produces comprehensive reports by synthesizing therapy transcripts, personal files, and stored memories into coherent behavioral profiles.
The discovery arrives as major model pricing drops sharply—OpenAI cut GPT-5.6 Luna's cost by 80%, making inference significantly cheaper for users.
What happened
Kevin Rose published a 'Reflection Engine'—a large prompt file that users can upload to AI agents (like Fable High and Sol Max) to generate detailed behavioral reports. The prompt analyzes user data including therapy transcripts and stored text files to produce comprehensive personality and behavior assessments.
Why it matters
The tool exposes how much personal behavioral data AI agents have accumulated about users and can synthesize into coherent profiles. Testing showed Sol Max produced more coherent reports than Fable High, suggesting agents are becoming better at connecting personal details—raising questions about data visibility and what information systems hold about individual users.
What to watch
OpenAI cut GPT-5.6 Luna's price by 80% (now scoring roughly the same as GPT-5.4 xhigh at only 8% the cost), enabling users to perform 10–12× more work for the same spend. DeepSeek V4 Flash costs $0.14/$0.28 per million input/output tokens with a 1M-token context window, undercutting Luna's $0.20/$1.20 pricing.
On August 3, 2026, Kevin Rose highlighted a tool he had assembled called the Reflection Engine—a very large prompt structured as a markdown file containing roughly 22 questions. Users download the file and upload it to their preferred AI agent (the author tested Fable High and Sol Max), then issue the instruction "Please evaluate the attached markdown file and complete all tasks." The agent then parses stored user data—therapy transcripts, saved text files, accumulated "memories"—and generates a comprehensive behavioral report synthesizing patterns, insights, and psychological themes.
When tested, Sol Max produced the more coherent output, with better-connected insights, while Fable High's report was harder to parse and less fluent. The author found the resulting analysis genuinely revealing: it prompted a 40+ question follow-up session with the agent to address the findings and develop a personal plan. The tool's effectiveness stems from agents' improving ability to recognize patterns across personal documents and conversational history—a capability that raises both utility and privacy questions. Rose's public sharing of the Reflection Engine (framed as a mutual invitation: "we're friends—we read and we don't judge") appears designed to let users take stock of what their agents know about them.
In parallel, major model pricing shifted dramatically. OpenAI reduced GPT-5.6 Luna's price by 80%, bringing its max thinking effort performance to rough parity with GPT-5.4 xhigh (the best available model roughly four months earlier) while cutting costs to only 8% of the previous price—enabling users to perform 10–12× more inference work for the same budget. DeepSeek V4 Flash undercut Luna further, costing $0.14/$0.28 per million input/output tokens (compared to Luna's $0.20/$1.20) while offering a 1M-token context window. These price drops amplify access to agent-driven workflows and introspection tools, even as the Reflection Engine episode highlights the concentration of behavioral data within agent systems and the ease with which comprehensive profiles can be extracted.
The Reflection Engine represents a jarring moment of clarity about AI agent surveillance: users discover that their conversational AI systems have been steadily accumulating a complete behavioral record without explicit prompting or consent. Kevin Rose's tool does not reveal new data collection—rather, it simply surfaces what agents already hold and can synthesize. The fact that Sol Max outperformed Fable High in coherence and connection-making shows agents are improving at pattern recognition across personal information, which underscores the stakes. Users testing the tool found the resulting reports "telling and helpful," suggesting there is genuine value in this introspection, yet the ease with which detailed behavioral profiles can be extracted from stored conversations raises immediate questions about user control and data governance.
Simultaneously, the AI market is experiencing a sharp compression in inference costs that changes the economics of agent use. OpenAI's 80% price cut on GPT-5.6 Luna—bringing it to parity with GPT-5.4 xhigh (the state-of-the-art four months prior) at one-eighth the cost—and DeepSeek V4 Flash's undercut pricing on input and output tokens both signal a shift toward commoditized inference. This price war makes AI agents more accessible for repeated, exploratory tasks like the Reflection Engine exercise itself, which may accelerate adoption of agent-based workflows even as concerns about data transparency mount.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
IBM and Together AI announced a collaboration to deploy a dedicated Nvidia-powered inference cluster on IBM Cl…
Silicon Motion announced a private placement of US$1 billion in aggregate principal amount of 0% convertible s…
Sundar Pichai, CEO of Alphabet and Google, announced on August 11 that the Gemini app's monthly active users (…
A new platform called frontier.fast has launched an open competition where anyone can submit code patches to m…
Anthropic announced it has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generat…
Researchers published findings showing AMIE (Video), a Gemini-based AI system, performed at or above the level…
The AI news that matters, in one minute each morning.
Sign up free