Anthropic’s Leaked KAIROS Code Unveils the New Claude Claw – Karpathy Reacts

When the Anthropic community examined the 510,000‑line Claude Code source that was unintentionally released, researcher Ole Lehmann uncovered a hidden module named KAIROS , which appears to be Anthropic’s internal answer to OpenClaw.

What KAIROS Does

Claude Code is evolving at light speed into Claude Claw.

KAIROS runs continuously in the background, receiving a heartbeat every few seconds. When the heartbeat triggers, it evaluates a short prompt roughly translated as “Wake up, see if there’s anything worth doing now?” Based on this assessment it can autonomously fix code bugs, reply to messages, update files, or execute tasks without the user issuing an explicit command.

Three exclusive capabilities are added on top of what Claude Code already offers:

Push notifications: KAIROS can send alerts to a phone or computer even when no terminal is open.

File delivery: Generated content is automatically delivered to the user without a request.

PR subscription: It watches a GitHub repository and reacts instantly to code changes.

These functions mirror OpenClaw’s core strengths—proactivity, personalization, and Skills—while integrating them into a single, out‑of‑the‑box product.

Personalization and Its Cost

KAIROS keeps a detailed daily log of everything it sees, decides, and does. The log is append‑only and cannot be deleted, so the longer the agent runs, the richer its context becomes. However, this continuous context growth inflates token usage dramatically; a simple "hi" in the morning can already consume tens of thousands of tokens.

Anthropic appears aware of this pain point and has introduced an "autoDream" routine that runs at night, consolidating the day’s experiences and reorganizing memory to mitigate context bloat.

Karpathy’s Prediction Comes True

In February, Andrej Karpathy predicted that a new AI layer—"Claw"—would follow Chat and Code, offering higher autonomy and proactive behavior. The KAIROS discovery confirms this vision: the agent acts before being prompted, then seeks user guidance only after completing work.

Examples cited include KAIROS automatically restarting a crashed server while the user sleeps and replying to a late‑night customer complaint before the user awakens.

Remaining Challenges and Open Questions

Despite its capabilities, KAIROS still suffers from heavy token consumption, especially when used with OpenClaw’s extensive context. Optimizing token efficiency is identified as a priority before a formal release.

The article also muses on the naming of this "lobster" agent, noting that the original project was called "Clawdbot" before being renamed under pressure from Anthropic, and wonders whether the final product will retain that name.

Reference links:

[1] https://x.com/karpathy/status/2039057005802082814<br/>[2] https://x.com/itsolelehmann/status/2039018963611627545