Good morning. Here is what matters in AI today, and how to put it to work.
Anthropic's own Claude Code workflow reveals 90% AI-written code, while practitioners find the agent's internals matter more than Anthropic admits.
~3 min read · last 12 hours
In today's issue
01
Anthropic: no mental model needed for Claude Code. Practitioners disagree.
02
Claude Code shares each background agent's full transcript with the orchestrating model
03
Instagram's Adam Mosseri: negativity around AI 'is going to get worse'
04
Death Clock promises to predict your lifespan. One user found it draining, not extending.
Main story
Anthropic: no mental model needed for Claude Code. Practitioners disagree.
Anthropic published its own Claude Code review workflow, noting the model now writes roughly 90% of its code, and argued teams do not need to understand Claude's internals to use it well. Practitioners are actively stress-testing that claim.
Why it matters: If your team is adopting Claude Code for production work, this debate directly shapes how much process and oversight you need to design around it.
What to watch next: Watch whether Anthropic publishes formal guidance on transcript handling and context isolation in multi-agent Claude Code setups, that documentation (or its absence) will tell us how seriously they take the production concerns practitioners are already surfacing.
Two separate looks at Claude Code this week reveal the same tension: Anthropic is telling teams to stop worrying about how the model thinks, while practitioners and researchers are finding that its internal plumbing (transcript sharing between background agents, 90% AI-written code) demands exactly the kind of mental model Anthropic says is unnecessary.
Share of Claude Code's own codebase now written by Claude, per Anthropic · InfoQ
Watch · On the feeds
Serving World Models: Latency vs Throughput with NVIDIA Cosmos 3 Super on Nebius | Cosmos Labs
NVIDIA Developer
API credits, now in ChatGPT Pro and Business plans
OpenAI
The Signal
The Claude Code disclosures this week mark a turning point for agentic development: when a model writes 90% of its own codebase and silently shares full agent transcripts, the "just use it, don't model it" advice breaks down fast. At the same time, the trust backlash Mosseri describes is not abstract sentiment, it is a product risk that compounds every time an AI app makes a personal promise it cannot keep. We see both threads pointing at the same lesson: opacity is a liability, whether inside your agent architecture or in how you communicate AI capabilities to users.
All the best, the KYFEX team
Quick hits
Agentic coding: what Claude Code actually does under the hood
Claude Code shares each background agent's full transcript with the orchestrating model
A technical finding shows Claude Code regularly passes every background agent's complete transcript to the main model, with real implications for context window consumption, cost, and information leakage between tasks.
Why it matters: Teams running multi-agent Claude Code workflows should audit what data flows between agents before scaling, both for cost control and to avoid unintended context bleed.
AI's trust deficit: from graduation stages to death-prediction apps
Two very different stories this week both point at the same underlying signal: public skepticism of AI is broad, personal, and growing, and the products most likely to deepen it are the ones that make high-stakes promises about individual lives.
Instagram's Adam Mosseri: negativity around AI 'is going to get worse'
Mosseri acknowledged in an interview that widespread dislike of AI is not limited to creatives or students, and predicted the backlash will intensify before it eases.
Why it matters: Product and engineering leaders should factor growing public distrust into rollout strategies, especially for consumer-facing features where AI is visible and personal.
Death Clock promises to predict your lifespan. One user found it draining, not extending.
The Death Clock app uses personal health data to forecast when you will die and nudge healthier behaviour, but a month-long test found the experience anxiety-inducing rather than empowering.
Why it matters: High-stakes personal AI predictions carry real psychological risk, and teams building health or wellness AI should weigh user wellbeing outcomes as seriously as engagement metrics.
Audit an AI agent workflow for unintended data sharing
I am running a multi-agent AI coding workflow. For each agent in the list below, help me identify: (1) what data it receives as input, (2) what data it produces as output, (3) whether any of that data could include sensitive context from another agent's session, and (4) what the minimum necessary context for each agent actually is. Flag any step where context bleed or over-sharing is likely. Agents: [paste your agent list here].
Why it helps: Given today's finding that Claude Code shares full background agent transcripts with the main model, running this audit now can catch cost and privacy surprises before they reach production.
Before you ship it
The risk
Multi-agent systems that share full conversation transcripts between agents can inadvertently expose user data, proprietary code, or credentials that were scoped to one task but become visible to others.
Do this
Define explicit context boundaries for each agent at design time: specify exactly which fields or transcript segments each agent is permitted to receive, and enforce those boundaries in your orchestration layer before deploying at scale.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
