
AevumOS by I2FLabs Vietnam: an autonomous multi-agent OS bringing AI squads on Gemini into the IDE
Vũ Khôi
Tác giả tại Wotech Electronics Blog

AevumOS by I2FLabs Vietnam: an autonomous multi-agent OS bringing AI squads on Gemini into the IDE
Instead of the "one-to-one" model — open a chat box, type a prompt, wait for a response, then copy the result — AevumOS organizes multiple AI agents into a "squad" with specialized roles, cross-review, and long-term memory, attached directly to the developer's existing IDE. According to a post by Nguyễn Hải in the Vibe Coding Chua group on 15 November 2026, it is a product of I2FLabs Vietnam (Innovation To Future Lab), introduced as a step from "chat tool" to Autonomous Multi-Agent OS, combined with new-generation Gemini models. This article presents the features as described in the original post — a promotional announcement — and Wotech has not independently verified these claims.
Quick answer: According to the post, AevumOS by I2FLabs Vietnam is not a new IDE but a standalone MCP server and "external brain" that attaches to Google Antigravity IDE, VS Code, or Cursor. The system organizes AI agents into a squad with specialized personas: the orchestrator agent Anton runs the data flow, chief architect Vidus reviews security vulnerabilities and guards the [GATE] review checkpoint, while An directly applies fixes and runs automated tests — no line of code is approved until it has passed the whole team's review. The features cited include the PBCE anti-context-drift mechanism, the Living Memory Graph for long-term memory, a CoW Sandbox virtualized environment on RAM, activeFiles file-scope locking, and a workspace jail barrier. Combined with new Gemini models, the post promotes three advantages: multiple Gemini sessions running in parallel with opposing roles, context-window optimization for faster responses and significant token savings, and full automation from idea to execution via the MCP protocol. First-time registered accounts get one month of the Aevum Pro plan free, according to the post.
An Autonomous Multi-Agent OS is how the post describes AevumOS: a coordination layer that turns multiple AI agents into an engineering organization with assigned roles, cross-review, and shared memory, instead of a single chat session. MCP (Model Context Protocol) is an open protocol standard that lets AI models connect to external tools and data sources; according to the post, AevumOS uses MCP to turn the agents' discussion and approval process directly into precise file-editing operations in the IDE.
What problem does AevumOS aim to solve?
According to the post, the way most of us work with AI still stops at the "one-to-one" model: open a chat box, type a prompt, wait for a response, then copy the result. But as technical problems grow more complex, a single AI session easily falls into context drift: the longer the conversation, the more it forgets the original architecture, makes sprawling edits on its own, and loses control over the real codebase.
That is why I2FLabs Vietnam was founded and built AevumOS, according to the post — to solve the core problem of the AI era: how can AI agents remember long-term, coordinate as a disciplined team, and operate safely on real systems?
Instead of chasing short-lived AI wrapper applications, I2FLabs Vietnam focuses on deep research into system architecture and artificial cognition, with research directions cited including the PBCE (Persona-Bounded Context Execution) anti-context-drift mechanism, spiking neural graphs, and the Organismic DDD (domain-driven design) programming model.
What is AevumOS, and where does it attach?
According to the post, AevumOS is not a new IDE, but a standalone MCP server and External Brain attached directly to the developer's existing working environment, such as Google Antigravity IDE, VS Code, or Cursor.
Instead of forcing one AI to "do everything", AevumOS organizes agents into a squad with specialized personas and a strict cross-control process. The core difference, according to the post, is this: instead of a single linear thought stream that easily overloads, the system orchestrates multiple agents taking on different roles on the same problem.

How does the AI squad work?
In a working session on AevumOS, as described in the post, the personas coordinate as follows:
| Persona | Role |
|---|---|
| Anton | Orchestrator agent, runs the data flow in the working session |
| Vidus | Chief architect, reviews security vulnerabilities and guards the [GATE] checkpoint |
| An | Directly applies fixes and runs automated tests |
The mechanism is called Role Assignment & Cross-Review (Cross-Session Orchestration): no line of code is approved until it has passed the whole team's review. One agent specializes in "attacking" to find architectural flaws, while another specializes in optimizing performance and writing test cases — opposing roles that check each other, rather than a single AI deciding everything alone.

How are long-term memory and safety handled?
Two features the post emphasizes are memory and safety:
- Long-term Memory (Living Memory Graph): Every architecture decision, codebase convention, and project history is connected into a living knowledge graph, helping the AI understand the project across many weeks of continuous work — eliminating the "amnesia syndrome" after each closed tab, in the post's words.
- Discipline & Safety: All command-line experiments run in a CoW Sandbox virtualized environment on RAM, with an activeFiles file-scope locking mechanism and a workspace jail barrier, ensuring the AI never breaks the original codebase structure.
These are the post author's claims about the product; Wotech has not tested or independently verified these mechanisms.
What advantages are promoted for new Gemini models on AevumOS?
According to the post, the newest generations of Gemini models bring lightning-fast reasoning, accurate multi-step tool use, and a huge context window. Placing that "reasoning engine" into the AevumOS frame, the post cites three advantages:
- Multiplying parallel thinking capacity: Activate multiple Gemini sessions simultaneously with opposing roles — one side specializing in attacking to find architectural flaws, the other in optimizing performance and writing test cases.
- Context-window optimization: Thanks to AevumOS's smart context-filtering mechanism, the Gemini model receives only the most essential knowledge fragments for each task, leading to faster responses, sharper reasoning, and significant token-cost savings — according to the post.
- Full automation from idea to execution: Through the MCP protocol, the entire discussion, review, and approval process of the Gemini agents inside AevumOS is converted directly into precise file-editing operations right in the IDE.
The post's closing message: the future of software development lies not in how fast you type prompts, but in how you build and orchestrate an ecosystem of intelligent agents.
What should readers note before trying AevumOS?
A few points readers should note, based on the post itself:
- This is a promotional announcement by I2FLabs Vietnam about its own product. All features and advantages cited are the author's claims, not independently verified by Wotech.
- Trial offer: According to the post, AevumOS's outreach campaign is underway, with first-time registered accounts getting one month of the Aevum Pro plan completely free. The registration link was left by the author in the original post's comments.
- No independent evaluation: The post includes no measurements, comparisons with other multi-agent systems, or third-party reviews — all of which would be needed before concluding anything about the system's real effectiveness.
Frequently asked questions
1. What is AevumOS? According to Nguyễn Hải's post, AevumOS is a standalone MCP server and "external brain" from I2FLabs Vietnam, turning AI into a multi-agent engineering organization (squad) attached to existing IDEs such as VS Code, Cursor, or Google Antigravity IDE.
2. Who is in the AevumOS AI squad? According to the post: Anton orchestrates the data flow, Vidus is the chief architect reviewing security and guarding the [GATE] checkpoint, and An directly applies fixes and runs automated tests. No line of code is approved until it has passed the whole team's review.
3. How does AevumOS handle context drift and memory? The post cites the PBCE (Persona-Bounded Context Execution) anti-context-drift mechanism and the Living Memory Graph — a living knowledge graph storing architecture decisions, codebase conventions, and project history across many weeks.
4. How does AevumOS keep the codebase safe? According to the post: command-line experiments run in a CoW Sandbox on RAM, with activeFiles file-scope locking and a workspace jail barrier so the AI cannot break the original codebase structure.
5. Is trying AevumOS free? According to the post, first-time registered accounts get one month of the Aevum Pro plan free; the registration link is in the original post's comments. Wotech has not verified this offer.
This article was rewritten from a post by Nguyễn Hải in the Vibe Coding Chua group on 15 November 2026, introducing AevumOS by I2FLabs Vietnam. It is a promotional announcement; all features, mechanisms, and advantages cited are the author's claims, not independently verified by Wotech.
Bài viết liên quan

16 tháng 11, 2026 · 9 phut doc
'After Watching This, Your Brain Will Not Be the Same': 4 Lessons on Neuroplasticity from Lara Boyd
A post by Dung Lê in a Facebook group on 14 November 2026 recounts the experience of watching a TEDx talk by brain researcher Lara Boyd — and how that talk changed the way the author understands lear…

16 tháng 11, 2026 · 7 phut doc
Breakscale: Don't Wait for Your App to Crash to Learn Where It Breaks — Break It First
You vibecode a web app, test it with 10 users and it runs fine, then run ads the next day, traffic pours in, and it crashes — and you have no idea where it broke or what to fix. That is the problem B…

16 tháng 11, 2026 · 8 phut doc
IT Students Blame AI for Unemployment: 'AI Isn't Stealing Jobs — Laziness and Delusion Are Eliminating You'
An anonymous post in the IT for Beginners Facebook group on the morning of 16 November 2026 is drawing attention with a counter-current argument: IT students who graduate unemployed are not victims o…