Public: Manus general autonomous agent

Architectural notes

Per the 2025 post "Context Engineering for AI Agents" by co-founder Yichao
"Peak" Ji: KV-cache hit rate is treated as the single most important
production metric (about 100:1 input-to-output token ratio), with stable
append-only prompt prefixes; tool availability is managed by a context-aware
state machine that masks logits during decoding instead of removing tool
definitions mid-task.

The same post documents the file system as "the ultimate context" --
unlimited, persistent, agent-operable memory (mapped here as memory: local),
a recited todo.md file to bias attention toward objectives, and a per-session
virtual machine sandbox. In a March 2025 response to a system-prompt leak
(per The Decoder), Ji confirmed Manus was built on Anthropic Claude 3.5
Sonnet plus fine-tuned Qwen variants and the open-source browser_use
project; those exact model versions are not in the catalog and current
versions are not publicly documented, so no model component is listed. NOT
publicly documented: sandbox infrastructure vendor, the in-house
orchestration framework, and eval/observability tooling.