Home
Aug 9, 2026
View All

Claude Code Auto Mode, Notion Acquires ZeroEntropy, and Pieter Levels Breaks Up with Claude

The Short Version#

Anthropic quietly made auto mode the default in Claude Code, Notion just acquired an AI model company to build task-specific intelligence in-house, and Pieter Levels is publicly done with Claude after fighting its verbosity problem for weeks — three signals that together say something interesting about where AI tool friction lives right now.

Anthropic / Claude Code - Auto Mode Is Now the Default#

Source: https://simonwillison.net/2026/Aug/8/auto-mode/#atom-everything Credibility: High (Simon Willison reporting on a shipped Anthropic product change, corroborated by the Anthropic changelog)

What happened: Anthropic flipped auto mode to the default for Claude Code on Pro, Max, and Team plans. Previously users had to opt in. This is a meaningful shift — it signals Anthropic's confidence that the autonomous coding agent behavior is reliable enough to be the default experience rather than an advanced setting users need to consciously enable.

Key product details:

  • Auto mode lets Claude Code operate more independently — planning, executing, and iterating without waiting for per-step user approval
  • The change applies across Pro, Max, and Team plans (not just enterprise)
  • Simon Willison flagged this as significant: making agentic behavior the default rather than opt-in is a threshold moment for the product
  • Users can still adjust autonomy levels, but they now have to opt down rather than opt in

Why it matters for PMs: The default-setting decision is a product philosophy statement. When you make autonomous behavior the default, you're betting that the marginal user's upside (faster output) outweighs the risk of unexpected actions. For any PM building agentic features, this is the tension to design around: how much autonomy do you default to, and what's your recovery path when the agent does something unexpected? Anthropic making this call at the plan level — not just for enterprise — tells you they've crossed a confidence threshold worth watching. It also directly connects to the open question about how much autonomy users are comfortable giving AI agents.

Critical questions:

  • What's Anthropic's rollback plan if auto mode causes a meaningful spike in user errors or complaints?
  • Is this change driven by user research showing people prefer auto mode, or by internal metrics showing it produces better output?
  • Does this change affect how Claude Code is positioned vs. Cursor and Windsurf, which have their own agent autonomy settings?
  • What does "auto mode" actually do differently in practice — is it just skipping confirmation dialogs, or is it making more substantive scope decisions?

Action you could take today: If your team uses Claude Code, flip to auto mode and run a session on a real task. Note where it makes scope decisions you'd want to override. That's your map of where user control still needs to live.

Notion - ZeroEntropy Acquisition#

Source: https://www.notion.com/blog/zeroentropy-is-joining-notion Credibility: High (first-party announcement from Notion's official blog)

What happened: Notion acquired ZeroEntropy, a company that builds "efficient, task-specific models made for knowledge work." This is Notion's first known AI model acquisition and signals a direct move toward owning the intelligence layer rather than purely relying on OpenAI, Anthropic, or other foundation model providers.

Key product details:

  • ZeroEntropy specializes in smaller, task-specific models — not general-purpose foundation models
  • The framing "efficient, task-specific models made for knowledge work" points toward Notion building models tuned specifically for things like summarization, Q&A over docs, writing assistance, and structured data extraction
  • This is a build move in the build/buy calculus — Notion deciding that the intelligence powering their AI features should be proprietary, not rented
  • The announcement was written by Sarah Sachs, suggesting a product/comms-led positioning

Why it matters for PMs: Notion just gave a concrete answer to one of the most live build/buy questions in product right now: when does it make sense to own your model layer? Their answer seems to be: when you have enough proprietary data (millions of user documents, workflows, and knowledge graphs) and a specific enough use case (knowledge work) that a task-specific model outperforms a general one. This is the same bet Stripe made with their internal models for fraud. The pattern to watch: as foundation models commoditize, the moat shifts to who has the best fine-tuning data and the best task-specific implementations. Notion just bought a team to build that.

Critical questions:

  • Does ZeroEntropy's model approach mean Notion plans to reduce its OpenAI API spend, or is this additive — specialized models for specific tasks alongside existing LLM usage?
  • What happens to Notion's existing AI features during the integration period? Users will notice if quality dips.
  • Task-specific models can be faster and cheaper, but they also have narrower capabilities. Where does Notion draw the line between "ZeroEntropy handles this" and "route to a frontier model"?
  • How does this change Notion's competitive position against tools that are doubling down on OpenAI integration (e.g., Microsoft/Copilot)?

Action you could take today: Pull up your AI feature roadmap and map each feature to its model dependency. Which ones could benefit from a task-specific model tuned on your own data? That's the ZeroEntropy bet applied to your context.

Pieter Levels - Breaking Up With Claude (And What It Reveals)#

Source: https://levels.io/deleting-claude + https://levels.io/stop-claude-from-getting-unintelligible Credibility: High (first-person account from a prolific builder-on-AI; Pieter Levels is one of the most visible practitioners shipping real products with AI tools)

What happened: Pieter Levels published two posts in one day: one announcing he's deleting Claude, another documenting "the prompt that finally stopped Claude from getting more and more unintelligible every day." He's been publicly frustrated for weeks with Claude's output quality degrading — becoming verbose, jargon-dense, and harder to use — and finally hit a wall. He found a system prompt workaround that fixes the behavior, then apparently decided that a tool requiring a workaround to behave reasonably isn't worth keeping around.

Key patterns:

  • Levels had been fighting Claude verbosity for weeks — this isn't a snap reaction, it's a documented pattern
  • He shared the actual prompt fix before deleting the tool, which is a practitioner gift: real signal about what Claude's default behavior does wrong
  • The "deleting" framing is notable: this is a high-frequency AI user making a decisive switch, not casual complaining
  • His frustration connects to a broader pattern he's been writing about: AI tools that start helpful and degrade over sessions or over time

Why it matters for PMs: Pieter Levels is an accelerated version of your power user. When he publicly documents a tool's failure mode and walks away, that's a retention signal worth analyzing. The specific failure here — output that becomes "unintelligible" over time, requiring a prompt workaround — points to a UX problem that no user should have to solve themselves. The PM question isn't "is Levels being fair to Anthropic?" It's: what failure modes in your AI features require users to develop workarounds? Those are the gaps your product team should own. Users who find the workaround stay. Users who don't, delete the app.

Critical questions:

  • Is Claude's verbosity problem a model-level issue (how Claude is trained) or a product-level issue (how the Claude interface is configured for different contexts)?
  • Levels found a prompt fix. Does Anthropic know this is a widespread problem? Is there telemetry that would surface it?
  • If a power user is hitting this wall, what's the silent majority of regular users experiencing?
  • What's the right product intervention — better defaults, user-controlled verbosity settings, or model-level fixes?

Action you could take today: Read Levels' prompt fix and test whether the failure mode he describes shows up in your own Claude usage. If you're building on the Claude API, check whether your system prompt is doing the work his workaround does — or whether your users are hitting the same degradation without knowing how to fix it.

Quick Hits#

The Thread#

The friction is in the defaults. Three stories this week point at the same thing: Claude Code making auto mode the default, Notion acquiring a model company to own task-specific behavior, and Pieter Levels needing a custom prompt to get Claude to behave reasonably. The underlying pattern is that AI tools are maturing past "does it work?" toward "does it work correctly without me having to configure it?" That's a product quality bar most tools haven't hit yet — and the ones that get there first will own retention.

Sit With This#

Pieter Levels found a prompt that fixed Claude's verbosity problem. He published it publicly, then deleted the app anyway — because a tool that requires a workaround to behave reasonably isn't a tool he wants to maintain.

For your AI product: What failure modes in your AI features are your users currently solving with workarounds? How would you even know if they were? And if a power user published the fix before churning, what would that tell you about where your product team's ownership needs to be?