Always-on runtime
KAIROS is presented as a system that can remain active across multi-step tasks instead of responding one turn at a time.
Kairos Files.
A reconstruction of the leak
The March 31, 2026 source-map leak suggested that KAIROS was not a simple chat feature, but a longer-running reasoning system built around recursive planning, persistent state, and delegated agent work.
What leaked
KAIROS is presented as a system that can remain active across multi-step tasks instead of responding one turn at a time.
The leak describes checkpointed state that can be saved, resumed later, and carried across sessions instead of living inside one context window.
KAIROS is described as a coordination layer that assigns specialized work to sub-agents and then combines their output.
The public response was a patch after a packaging error exposed internal logic that was not intended for release.
Timeline
Anthropic says it patched the exposure. People are still analyzing the leaked code, but the original packaging mistake appears to have been fixed.
Evidence blocks
Backed by the article’s checkpointing and persistence framing.
Derived from the article’s planner and commander-system framing.
Pulled from the article’s incident and response summary.
The Reddit thread expands the leak into a wider mythology, but the article grounds the core story in three clearer mechanisms: recursive planning, persistent checkpoints, and multi-agent delegation.
That makes it easier to separate mood from evidence. The rumor layer still contributes atmosphere, but the article gives the page a firmer technical center.
Final state
system transcript reconstructed from the exposed interface layer.