The Hidden Prologue of AI Awakening: How Manus and Cursor Split the Fog of the AI Era with Cognitive Insight

> Structured English rendering of a Chinese forum post. Numbers and claims are as stated by the original author.

Key points

• In April 2026, Meta agreed to acquire Manus for $2 billion and Elon Musk (SpaceX/xAI) offered ~$60 billion for Cursor. Common Chinese internet commentary dismissed both as "wrappers" over others' models and the bids as impulsive. The author argues this is a directional misjudgment: both teams were cognitively ahead of the industry by at least half a year to a year.

Manus: refusing role-playing

• Since its March 2025 launch, Manus was criticized as a wrapper using Claude and Qwen without a self-trained model (MIT PhD Qin Zengyi: excellent product, not a technical breakthrough).
• Core insight: no "hat wearing." 2023–early 2025 multi-agent systems like MetaGPT assigned human roles (product manager, architect, PM, engineer, QA) in serial pipelines. But LLMs are generalists; forcing roles only restricts capability. Manus's *wide research* mechanism instead spawns independent full-capability Manus instances, each with its own context window in cloud VM sandboxes, splitting only at the task level.
• By 2026 this became industry consensus: OpenAI Codex (Plan/Spec Mode), Anthropic Claude Code (orchestrator-worker), Cursor (Planner-Worker-Judge) all converge on functional division of labor, not human job titles.
• Manus was also first (March 2025) to deliver an end-to-end chain—autonomous search, code generation, data visualization—and to let users deploy generated apps with one click, with an API so the generated software itself can keep calling Manus's AI. This reframed AI from a one-shot productivity tool to infrastructure for *User Generated Software*; its waitlist passed 2 million after the demo.
• Results: $100M ARR in 8 months, 1,470 trillion tokens processed, 80M+ virtual machines created, GAIA Level 3 score 57.7% vs OpenAI Deep Research's 47.6%.
• Its July 2025 context engineering post (keep prefix stable; make context append-only; mask tools, don't remove them) was widely adopted across harness engineering and validated building agents on top of frontier models' in-context learning rather than training an end-to-end agentic model from open weights—a minority view then, mainstream now.
• Answering objections: (1) Manus's cloud-sandbox form is one generation behind local terminal agents, but acquisitions buy the team's cognition, engineering, users, and infrastructure—Meta had already integrated Manus agent capabilities into Ads Manager by Feb 2026. (2) If it were truly a content-less wrapper, China's NDRC would not have used the Foreign Investment Safety Review regime's first-ever "prohibition plus revocation" in five years to block the deal in April 2026, deeming its team, R&D, training data, and IP national-security-relevant assets.

Cursor: the only third party that forged its own models

• Coding agents hinge on high-frequency tool calls where latency directly determines UX. Cursor judged early that self-trained models were a product necessity—unlike competitors (Cline: open-source VS Code plugin with third-party models; Trae (ByteDance): Claude 3.5 Sonnet + GPT-4o; Windsurf/Cognition: UI and workflow differentiation).
• Composer 1: built on an open-source MoE base with large-scale RL post-training in an agent harness simulating Cursor's production environment. In the author's experience it matched Sonnet 4.5 on ~90% of daily tasks while returning in seconds instead of minutes—the bottleneck is speed/cost, not capability ceiling.
• Composer 1.5: 20x RL compute, added thinking tokens and self-summarization—but diminishing returns (CursorBench +6.2).
• Composer 2: methodological shift—continued pretraining before RL, on the Kimi K2.5 base (confirmed by Moonshot), yielding +17.1 on CursorBench and stated Pareto optimality at significantly lower inference cost. This aligns with ICML 2025's "SFT Memorizes, RL Generalizes" and Moonshot's Kimi K2 report, which Cursor anticipated in product.
• Only first-party LLM providers (OpenAI Codex, Anthropic Claude Code) and Cursor run the complete four-stage pipeline (base selection → continued pretraining → RL post-training → product integration); Cursor is the only third-party startup to do so.

Harness engineering and spatial scaling

• Cursor's Feb 2026 *self-driving codebases* experiment used recursive Planner-Worker architecture with hundreds of parallel agents, peaking around 1,000 commits/hour and generating over a million lines of Rust—framing the question as how 10x compute yields 10x meaningful throughput.
• After critics found failing CI in public repos and alleged fraud, the fairer reading is that Cursor honestly documented failure modes (error-rate trade-offs, hallucinated dependencies, agents stepping on each other) in the original post; the experiment's code quality was indeed poor, but no competitor explored multi-agent throughput scaling at that scale. Cline and Trae were still optimizing single-agent interaction quality; Cursor had also shipped background agents and parallel task execution in the UI earlier.
• Fortune's March 2026 "Cursor's crossroads" piece debated not whether Cursor can compete with LLM providers, but how it keeps its lead after they entered—no real wrapper earns that framing.

The shared pattern

1. Cognitive leadership in their domains (Manus: generalist agents, user-generated software, context engineering; Cursor: full post-training pipeline, harness engineering, background agents)—ahead of peers by 6–12 months. 2. Results validated by top buyers: Meta's $2B and SpaceX/xAI's $60B price this judgment. 3. The "wrapper" label reflects an outdated taxonomy. By 2025, training-pipeline design, agent architecture, context engineering, and harness engineering matter as much as—or more than—self-trained models. Judging 2025 products with a 2023 framework systematically underestimates their depth. 4. Real challenges remain: the NDRC blocked the Manus acquisition, leaving its path uncertain; Cursor faces frontal competition from Claude Code and Codex, with core engineers beginning to move to xAI. Neither changes the fact that both teams rank in the industry's top tier for technical judgment in the agent AI era.

---

References (as cited in the original post) 1. Manus team. *Context Engineering: Principles for Agentic AI Systems.* July 2025. 2. Cursor official blog. *Composer 2: Post-Training for Interactive Programming.* 2026. 3. MetaGPT project docs. *Multi-Agent Framework with Role-Based Collaboration.* 2024. 4. GAIA Benchmark Consortium. *Level 3 Evaluation Report on Agentic Systems.* 2026. 5. ICML 2025 Proceedings. *SFT Memorizes, RL Generalizes: Implications for Post-Training.* 2025.