Manus and the Agent-First Product Approach

Manus leaders describe an agent-first product strategy centered on cloud-resident persistent agents, multi-model orchestration and heavy context engineering. - Founder Peak's GPT-3 experience changed his view on training vs. using large models. - Manus hired Peak as Chief Scientist; CEO Red is described as steady operational leader. - Manus abandoned a local AI-native browser and moved to a cloud 'Browser-in-the-Air' design. - Manus leverages external SOTA models instead of training base models to avoid long iteration cycles. - Manus uses multi-model routing: Anthropic for coding, Gemini for multimodal, OpenAI for reasoning as needed. - Manus emphasizes agent data: trajectories, tool use, trial-and-correct loops as unique product assets. - Manus restricts early access to control compute demand and later removed invites after capacity scaled. - Peak argues for 'pure-blood' agents that self-observe and self-correct instead of brittle rule patches. - Recommendations to model providers: compress context, favor tool-integrated incremental reasoning, handle interruptibility and improve recovery from errors. - Manus targets prosumers; pricing and DAU are secondary to Agentic Hours and compute sustainability.
Key insights
Manus chooses cloud-resident agents over local browser integration
Manus shifted from building a local AI-native browser to a cloud-based 'Browser-in-the-Air' where each user has a persistent cloud VM for long-running agents. The team found local integration caused resource contention and disrupted user workflows; cloud residency enables agents to maintain context, run 24/7, and operate independent toolchains. This architecture was inspired by Monica, a browser plugin that demonstrated seamless context capture without changing user habits. Manus uses persistent cloud agents to provide continuous background work, integrate multi-source context and avoid 'computer stealing' UX problems observed on local clients. (supporting_posts: [20028])
Manus refuses to train base models and instead orchestrates multiple leading models
Manus deliberately avoids building its own base models to prevent product iteration cycles from being locked to checkpoint timelines. The team treats training in-house base models as a 'model lottery' and instead purchases top-tier model access across providers—Anthropic for coding, Gemini for multimodal understanding, OpenAI for reasoning—becoming a top consumer of inference. This multi-model strategy gives Manus flexibility to pick best-fit models per task, influence provider roadmaps, and capture unique agent usage data (full trajectories and tool interactions) that are not available to model vendors. (supporting_posts: [20028])
Manus treats agent interaction traces as a proprietary product flywheel
Manus emphasizes that agent data differs from simple chat logs: it records entire tool-using trajectories—planning, API calls, code edits, self-corrections and user adjustments. Those trajectories form a distinct dataset enabling product improvements and a behavioral flywheel. Peak argued this agent-level feedback is richer than pairwise prompt-response data and supports learning about tool orchestration, error recovery and user preferences, which are central to improving long-running task performance and product stickiness. (supporting_posts: [20028])
Manus practices conservative rollout to manage exponential compute demand
Manus used an invite-only launch due to extreme backend compute demands for agentic reasoning. Agents consume orders-of-magnitude more tokens than chatbots (estimates of 100x–1000x), creating capacity pressure; cloud partners warned Manus that uncontrolled public access could saturate infrastructure. Manus prioritized servicing prosumers—users insensitive to cost but sensitive to correctness—focusing on Agentic Hours rather than DAU. When compute capacity stabilized, Manus rapidly removed invite constraints. (supporting_posts: [20028])
Product philosophy: build a digital 'person' not vertical tools
Manus differentiates between vertical 'tools' (e.g., Cursor) and a general-purpose digital 'person'—an agent with eyes, ears and hands that integrates across interfaces. Peak argues that vertical tools accelerate parts of workflows, but only a general agent can orchestrate cross-tool tasks (e.g., fetch data from GitHub, run tests, write docs and coordinate Slack). Manus calls its approach 'pure-blood Agent'—avoiding hard-coded workflows and instead relying on agent observation, reflection and iterative repair to handle edge cases. (supporting_posts: [20028])