Claude Sonnet and Jarvis Pattern: why AI agents might not need more than an operating system

An article published on Habr proposes radically simplifying the discussion about AI agents: instead of frameworks, graphs, orchestrators, and vector databases, the author proposes the formula LLM + operating system + files. The core idea is that modern models are already powerful enough to use terminals, APIs, and file structures almost as an engineer would, meaning the bottleneck isn't the amount of infrastructure around the agent but how its memory, access rights, and work environment are organized.

Egor Zinovyev, an IT architect, describes the Jarvis Pattern as a personal network agent attached to a specific specialist. Such an agent operates in its own container with access to infrastructure and acts on behalf of the person.

As an example, he presents his DevSecOps agent umax based on Claude Sonnet, which according to his account covers a complete set of specialized tasks: from RBAC configuration in Kubernetes and Vault operations to cluster deployment, Docker image scanning, CI/CD setup, and vulnerability analysis.

The main thesis is that the agent doesn't need a pre-assembled tool kit: if a Prometheus query is needed, it uses curl; if data transformation is required, it uses jq, sed, or awk; if no suitable utility exists, it writes one and adds it to the workflow.

Special emphasis is placed on memory. According to the author, it remains the truly unsolved part of agent architecture. He proposes dividing it into declarative, procedural, and episodic memory: facts, instructions, and experience. Particularly important is not only successful experience but also negative experience—knowledge about which paths have led to dead ends and why.

Instead of vector search and graph models, Zinovyev bets on the file system and markdown files as a natural route map: folders define categories, file names indicate direction, and index documents serve as entry points for diving into the necessary context. In parallel, a separate Memory Agent should work to analyze after each session what to save, what to update, and what to discard.

From this logic emerges a broader view of the human role. Jarvis Pattern isn't about fully autonomous AI but about augmenting a specific engineer: the person sets the task, verifies the result, and makes decisions under uncertainty, while the agent handles execution and routine work. The author believes such a model can transform hiring, as candidates can be evaluated not by abstract questions but by how they work with the agent on real cases. The article even cites enterprise market benchmarks: hiring cycles of 40–60 days, error costs representing tens of percent of annual salary, and a notable share of employees not lasting the first year.

Another practical conclusion concerns software: if the agent works through API and CLI, products without proper APIs will start losing regardless of how beautiful their interface is.

Although the text is presented as a manifesto, the author emphasizes he doesn't consider himself alone. As proof, he cites similar ideas from other engineers and products where the agent is already perceived as an operating environment for the specialist rather than a chat with buttons.

In this sense, Jarvis Pattern is neither a finished standard nor a new platform but an attempt to capture a shift: part of the industry is beginning to view the AI agent not as a separate application but as a management layer over existing infrastructure.

If this logic takes hold, the main debate around AI agents will shift from framework choice to memory design, access rights, and API-first tools. For teams, this means less architectural magic and higher demands on context quality; for specialists, it means growing value of deep domain knowledge, which the agent can scale but not replace.