Google introduced subagents to the Gemini CLI, transforming the traditional terminal into a multi-agent dispatch center.
Developers can now summon specialized expert agents directly from the terminal. They use a simple “@agent” syntax to delegate work. You create an agent, give it a specific job, and cause it to run in parallel with your primary coding session.
When an AI is fed too many files, it suffers from what Google accurately identifies as context rot. The model forgets previous instructions, starts hallucinating variables, and generally loses plot completely.
Engineers who work on legacy enterprise systems are familiar with this problem. You spend hours preparing an algorithm with architectural rules, only to watch it generate conflicting logic ten prompts later. Because context windows span millions of tokens, the noise drowns out the relevant data.
Subagents solve this problem by isolating the context windows. Each subagent receives exactly the data it needs to complete a high-volume task. The primary session stays quick and remains focused on the overall goal.
When a developer needs to refactor a database schema, they don’t ask the main AI to do it. You type “@db-agent” and unload the work. The subagent starts in its own isolated environment, reads the relevant files, runs code generation, and returns the finished product to the primary session. The main model never has to clog its memory with irrelevant backend logic while trying to help the developer write frontend UIs.
Microservices for AI
Command line interfaces have evolved little since the days of Bash and Zsh. Text goes in and text comes out. Now the CLI houses a complete virtual engineering department.
Development teams write simple Markdown files to customize these subagents. They define the agent’s expertise, its limitations, and its execution environment. They give him a persona and strict limitations. When the primary Gemini CLI encounters a problem that requires this specialized expertise, it automatically delegates the work or waits for the manual @agent trigger.
Security teams will also appreciate this modular approach. Markdown files provide a clean, readable audit trail. A security engineer can open the configuration file and see exactly what permissions a particular subagent has. This is consistent with the principle of least privilege; If an agent is configured to read only the UI components folder, it cannot accidentally rewrite the core database configuration.
Delegation takes place in parallel. An engineer can trigger three different subagents to complete three different background tasks at the same time. One agent updates API documentation, another writes unit tests for a new payment gateway, and a third scans the codebase for deprecated libraries. The human developer sits at the center of this web, orchestrating execution and reviewing pull requests.
An ecosystem moving towards orchestration
This week OpenAI updated its Agent SDK with native sandbox execution and a model-native harness. Their goal is the same: to help developers build secure, long-lasting agents that can traverse files and tools autonomously.
While OpenAI attacks the problem from the software development kit level, Google embeds the solution natively in the terminal.
The terminal remains the developer’s undisputed home. By embedding intelligent orchestration natively into the CLI, friction is avoided. This keeps engineers in a flow state without having to switch between web browser windows and code editors. Competing platforms must offer similar delegation protocols or risk obsolescence.
By decomposing complex engineering tasks into specialized subagents, you deliver software at scale. Deterministic execution replaces probabilistic estimates. They force the AI to use certain tools in a specific order, rather than hoping it deduces the right path from a vague instruction.
The hard work of software development takes a back seat and is carried out by specialized agents that run in parallel in the terminal. Human engineers define the boundaries, write the Markdown constraints, and approve the final pull requests.
See also: HubSpot just fixed one of its most frustrating developer issues
Want to learn more about AI and big data from industry leaders? Checkout AI and big data trade fair takes place in Amsterdam, California and London. The comprehensive event is part of TechEx and takes place alongside other leading technology events including the Cybersecurity and Cloud Exhibition. Click Here for more information.
The developer is supported by TechForge Media. Discover more upcoming enterprise technology events and webinars Here.
