Reverse-engineered from public sources and leaked artifacts" became "reconstructed entirely from public sources — the team's blog posts, conference interviews, leaked system prompts, and ecosystem case studies." The expansion does two things — it backs up the "reverse engineered" claim by naming the kinds of sources, and the word "entirely" preempts any concern about whether the doc relies on insider information. It doesn't.
Each layer in layers/ has its own folder containing:
- A deep-dive README explaining what the layer does, the design decisions behind it, and how data flows in and out.
- A phase-specific architecture diagram for that layer.
- An Open-Source Toolkit section listing the actual libraries and services you can use to build that layer yourself.
| # | Layer | What it does | Color in HLD |
|---|---|---|---|
| 00 | Event Stream | Append-only WAL — the system's source of truth | amber |
| 01 | Ingress / Router | Auth, rate limit, session, Firecracker microVM boot | pink |
| 02 | Planner Agent | Task decomposition, schema definition, sub-agent spawning | purple |
| 03 | Sub-Agent Pool | The centerpiece — context-isolated parallel workers | green |
| 04 | Executor Agent | The user-facing ReAct loop | cyan |
| 05 | Knowledge Manager | Async trajectory tailing and artifact persistence | orange |
| 06 | Tool Execution Layer | Browser, shell, filesystem, deploy subsystems | slate |
| 07 | Model Router | Task-class to LLM mapping with cache-aware routing | red |
| 08 | Context Engineering Principles | The eight cross-cutting rules that hold it together | red |
If you're new to the architecture, read in this exact order:
layers/00-event-stream/— start here because every other layer interacts with it.layers/08-context-engineering/— read the rules before the layers, because the rules dictate the layers' designs.layers/01-ingress-router/throughlayers/07-model-router/in numerical order.
If you're already familiar with multi-agent systems and want the new ideas first:
layers/03-sub-agent-pool/(the multi-agent protocol)layers/08-context-engineering/(the KV-cache mechanics)layers/06-tool-execution-layer/(the dispatch-by-prefix trick)
diagrams/architecture-hld.png(also.svg) — high-level architecture (full system)diagrams/low-level-systems.png(also.svg) — per-iteration ReAct loop + KV-cache state + tool dispatch tree
script/linkedin-video-script.md— 5-minute LinkedIn-style hot-take scriptscript/diagram-walkthrough.md— 6-minute screen-share narration with "point at this" cuesscript/talking-points.md— outline-style version if you prefer speaking from beats
All architectural claims are sourced from public material — see references/sources.md for the full bibliography. The big ones:
- Peak Ji — Context Engineering for AI Agents: Lessons from Building Manus
- Lance Martin × Peak Ji — webinar notes, Oct 2025
- E2B case study — How Manus uses E2B
- Leaked
manus_20250310system prompt + tool.json
Where a claim isn't explicitly documented (e.g., the CDP layering inside browser-use, the use of forkpty/execve for shell tools, the directory layout under /home/ubuntu/), it's drawn from standard practice for that kind of subsystem. Every layer README flags these inferences explicitly. If you find a documented contradiction, open an issue.
This repo is documentation, not running code. But every layer's Open-Source Toolkit section lists the libraries you'd actually wire together to build that layer. If you want to clone the architecture, walk the layers in build order: 00 → 01 → 04 → 06 → 02 → 03 → 05 → 07. Don't try to build the sub-agent pool before you have a stable event stream and a working tool dispatcher — that's the order in which Manus itself evolved over five rewrites.
If you find a documented contradiction with one of the primary sources, please open an issue. See CONTRIBUTING.md for what kinds of contributions are most useful.
This documentation is released under Creative Commons Attribution 4.0 International (CC BY 4.0) — you can copy, redistribute, adapt, and build on it (including commercially), as long as you give appropriate credit and link to the license.
Manus and the names of underlying providers (Anthropic, Google, OpenAI, E2B, etc.) are trademarks of their respective owners. This is independent documentation of a publicly described system and is not affiliated with, endorsed by, or sponsored by any of those parties.