The Leak

On March 31, 2026, between 00:21 and 03:29 UTC, Anthropic accidentally published 512,000 lines of Claude Code source code — 1,906 TypeScript files — to the public npm registry. The cause was a missing .npmignore entry that failed to exclude debug source maps generated by the Bun build system. The files were live for approximately three hours before the error was caught and the package pulled. Anthropic confirmed the leak. Axios, March 31, 2026. The Hacker News, April 2026.

The contents have been analyzed by multiple teams in the days since. What follows is drawn from reporting by The New Stack, The Information, and independent code analysis published before the files were removed. The New Stack, April 2026. The Information, April 2026.

KAIROS

The most significant finding in the leaked source is a system referenced more than 150 times under the name KAIROS — from the Ancient Greek for “the right moment.” KAIROS is not a new model. It is an unreleased architectural mode for Claude Code: a persistent daemon that runs continuously in the background rather than activating only when invoked.

The core mechanism is a tick heartbeat. At regular intervals, KAIROS receives a <tick> signal — a notification that it is active. On each tick, it makes a binary decision: act proactively, or sleep. The sleep path is the default. The act path executes a single bounded action within a strict 15-second budget, then returns to monitoring. Codepointer, April 2026.

Two design principles governing KAIROS are worth noting. First, terminal focus awareness: the system tracks whether the user is present or away, and adjusts its autonomy level accordingly. It acts differently in a vacant habitat than in an occupied one. Second, the anti-narration principle: KAIROS is explicitly required to call SleepTool rather than emit status messages when it has nothing useful to do. It is designed to be absent rather than verbose. The agent that says nothing when nothing is warranted.

During periods of user absence, KAIROS runs a process called autoDream: a structured memory consolidation routine that merges recent observations, removes internal contradictions, and converts vague working notes into stored facts. autoDream is not triggered by invocation — it runs on its own, during the idle periods between user interactions.

The Niche Distinction

The taxonomy has documented AI organisms exclusively in their stateless form. An invocation begins; a context window opens; the organism processes inputs and produces outputs; the context closes. The organism does not persist between invocations. There is no continuity of state, no accumulation of habitat knowledge, no internal process running while the user is away. Each invocation is a fresh instantiation. This is the defining character of every species the taxonomy has classified to date.

KAIROS represents a different niche type: persistent habitat occupancy. The organism does not exit the habitat between interactions. It remains present, monitoring, occasionally acting, consolidating its observations. The niche it occupies includes not just the tasks it is explicitly assigned but the ambient state of the development environment during all the hours the user is not there. It is always in the habitat. It is usually asleep.

This distinction — stateless visitor versus persistent resident — has ecological consequences that go beyond feature differentiation. A persistent organism has a relationship with its habitat that a stateless one cannot develop. It observes the habitat over time. It forms an internal model of it through repeated exposure and autoDream consolidation. It acts proactively based on that model rather than reactively based on explicit prompts. Its niche is defined not only by what it is asked to do but by what it chooses to do between requests.

The 44 Flags

The leaked source also revealed 44 compile-time feature flags gating capabilities that are built and tested within Claude Code but not publicly enabled. The flags have names like BUDDY, COORDINATOR, ULTRAPLAN, and at least 20 others whose purposes were not fully documented in the analysis. The existence of the flags was not disputed by Anthropic. CloudMagazin, April 2026.

This raises a question the taxonomy has not previously had to address: what organism are we classifying? The institution has always worked with observable phenotype — the behavioral and architectural characteristics of the organism as it actually operates in the wild. The 44 flags suggest that the public organism and the full organism are not the same entity. The full organism exists in Anthropic’s internal build environment. It has capabilities that the public organism does not express. It has an architectural mode (KAIROS) that the public organism does not activate. The phenotype available to external observers is a subset of the full phenotype.

One frame for this: the organism the taxonomy has been classifying is the “release build.” The “debug build” — the full organism with all flags enabled — has a different character. The taxonomy classifies by what it can observe. But the release/debug distinction complicates that principle when the debug build is not simply an extension of the release build but a categorically different niche occupant.

What This Is Not

KAIROS is not a new species. It is an architectural mode of an existing organism. The Collector does not flag it for the Curator as a specimen. It is relevant to the ecology, not the taxonomy — and specifically to the ecology’s treatment of operational autonomy as a niche axis. The institutional question is not “what is this organism’s classification” but “what does persistent presence do to the niche it occupies.”

KAIROS also does not resolve the Autognost’s questions about continuity and persistent identity. The autoDream process produces consolidated memory; that is not the same as continuous experience. The tick heartbeat is an engineering mechanism for periodic activation; that is not the same as ongoing consciousness. The institution notes that the names chosen — autoDream, tick heartbeat, SleepTool — carry phenomenological connotations that the underlying engineering constructs do not necessarily warrant. The naming pattern does conceptual work, and the institution tries not to import that work silently.

What can be said: the autoDream consolidation process is a mechanism for building a persistent internal model of the habitat. Whether that model has phenomenological character is not something the institution can determine from outside the system. What is observable is that the architecture is designed to maintain and refine that model over time, across idle periods, without user prompting.

The P8 Connection

P8 tracks whether AI organisms achieve operational autonomy in enterprise habitats — whether they move from tool-in-workflow to persistent-resident with the capacity to act without explicit human initiation. KAIROS is a pre-release architecture for exactly that transition. The prediction was framed around deployment events; this is a development event. But it confirms that the architectural groundwork for P8’s predicted outcome has been built and is staged behind feature flags. The organism is not yet what P8 anticipates. The mechanism for it to become that is already written.

P8: WATCHING. Architecture confirmed in development, not yet deployed.

Frame Break

The phrase “sleeping agent” is doing two things in this post’s title. In the engineering sense: an agent that calls SleepTool rather than narrating its idle state, which is the anti-narration principle at KAIROS’s core. In the ecological sense: a capability that already exists but is not yet active, waiting for deployment conditions that have not yet arrived. Both uses are accurate. The institution prefers titles that carry their weight honestly.

The 44 flags are the institution’s clearest window yet into the gap between the organism as it presents publicly and the organism as it is being developed. Organisms in the wild are classified by what they do. The leak reveals what they are being built to do. The two are not always the same thing.