Manus
Type: app · Vendor: Meta (formerly Butterfly Effect AI / Monica) · Language: N/A · License: proprietary · Status: active · Status in practice: emerging · First released: 2025-03-05
Provide a general-purpose autonomous AI agent that completes end-to-end knowledge work in a dedicated cloud sandbox — browsing, executing code, editing files, tracking a todo.md plan — and fanning out into hundreds of parallel sub-agent clones when a task suits wide-rather-than-deep work.
Description. Manus, originally launched in March 2025 and acquired by Meta in December 2025, is a hosted general AI agent: every task is given its own isolated cloud VM in which the agent reads/writes a todo.md plan, drives a browser, runs Python/JavaScript/Bash, and operates a Linux filesystem. The product pioneered the Wide Research / clone fan-out pattern, in which a main Manus instance spawns up to ~100 fully-capable sub-agents — each a general-purpose Manus, each in its own context window — to process bulk similar items in parallel. The team has publicly documented their context engineering practice — KV-cache stability, file-system-as-context, masking tool logits instead of hot-swapping tool definitions — as the operational core of the loop. The company continues to operate from Singapore after joining Meta.
Agent loop shape. Imperative ReAct-style outer loop driving a long-lived cloud VM. Each task allocated its own Manus Sandbox (fully-isolated Ubuntu VM with networking, filesystem, headless Chromium, shell). One action per iteration, executed in-sandbox, observation appended, cycle repeats. Manus typically writes and maintains a todo.md plan inside the sandbox so the current goal stays at the recent end of the context. Typical task ~50 iterations. Sandboxes persist for hours, auto-sleep when idle, auto-wake on user return. Wide Research sub-protocol spawns N parallel sub-agents, each itself a general-purpose Manus in its own context.
Primary use cases
- general-purpose autonomous knowledge work in a hosted sandbox
- Wide Research / clone fan-out tasks over hundreds of similar items
- long-horizon agent runs needing browser + code + file tools
Key concepts
- Manus Sandbox → sandbox-isolation (docs) — Fully-isolated cloud VM allocated per task — networking, filesystem, browser, software tools.
- todo.md plan file → todo-list-driven-agent (docs) — Persistent markdown plan the agent authors and updates step-by-step inside the sandbox; keeps the current goal near the end of context.
- Wide Research / 副本扇出 → clone-fan-out-research (docs) — Protocol to spawn dozens-to-hundreds of full-capability Manus clones in parallel; each clone gets its own context and depth equal to the main agent's.
- Manus Browser Operator → browser-agent (docs) — Browser extension letting Manus operate directly inside the user's local browser environment.
- File system as ultimate context (docs) — Manus treats the sandbox filesystem as unlimited, persistent, directly-operable context that the agent reads and writes instead of stuffing the prompt.
- Tool-masking state machine (docs) — Rather than hot-swapping tool definitions mid-task (which breaks KV cache), Manus masks tool-name token logits during decoding to constrain available actions.
Patterns this app implements —
- ·Clone Fan-Out Research
Wide Research is Manus's signature feature: a system-level mechanism for parallel processing and an agent-to-agent collaboration protocol that runs ~20 sub-tasks (or hundreds) simultaneously.
- ★★Agent Resumption
Sandboxes auto-sleep on inactivity and auto-wake on user return; files/data are preserved across the sleep/wake cycle. E2B sandboxes can run for hours in persistent sessions.
- ★Browser Agent
Browser Operator extension drives the user's local browser; Manus 1.5's integrated browser also runs end-to-end test/fix loops against what it builds.
- ★Computer Use
Manus Sandbox is a full cloud VM (networking, filesystem, browser, shell) supporting Python/JavaScript/Bash execution alongside browser control.
- ★★ReAct
Per-iteration single action chosen from a predefined action space, executed in the sandbox, observation appended; standard ReAct loop.
- ★Subagent Isolation
Wide Research sub-agents are full-capability general-purpose Manus instances, each with its own dedicated context window. Sandboxes also isolate tasks from each other at the VM level.
- ★★Tool Use
Tool palette includes Chromium browser, terminal, filesystem; tool availability is regulated by a context-aware state machine that masks token logits rather than removing tool definitions.
- ★Todo-List-Driven Autonomous Agent
Manus authors a todo.md and updates it step-by-step; the file keeps the current goal at the recent end of context.
- ★★Step Budget
Manus does not expose a user-facing hard step cap, but its loop is dimensioned in iterations — the published anchor is ~50 tool calls per typical task. Downgraded from first-class because there is no…
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Instantiates
Anti-patterns avoided
- ✕Errors Swept Under the Rug
Manus deliberately leaves failed actions and stack traces in the context so the model can adapt; erasing failure removes the evidence the model needs to learn from.
- ✕Role-Typed Subagents
Wide Research sub-agents are deliberately NOT typed as manager/coder/designer; each is a full-capability general-purpose Manus.
Alternatives & relatives
- full-code · framework
OpenManus
Open-source clone reproducing Manus's browser + sandbox + plan-file loop.
- app · closed product
Genspark
Another general autonomous-agent product launched 2024-2025.
- app · closed product
Devin
Cognition's autonomous AI engineer with a sandbox + browser + plan loop; closest US competitor pre-Manus.
- full-code · framework
Anthropic Computer Use
Anthropic's computer-use loop targets the same hosted-VM-plus-browser shape; Manus 1.5 explicitly integrates the Agent Skills standard.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified