Agents shift to subagents

A coding agent is starting to look less like one all-purpose bot and more like a manager that hands work to specialist bots. (developers.openai.com) (developers.googleblog.com) OpenAI’s Agents Software Development Kit docs now describe agents that can “hand off to other specialized agents,” while a separate orchestration guide says teams should split work only when a branch needs different instructions, tools, or policy. (developers.openai.com 1) (developers.openai.com 2) Google pushed the same pattern into Gemini Command Line Interface on April 15, 2026, saying subagents can take on “complex, repetitive, or high-volume tasks” in their own context windows while the main session stays focused. (developers.googleblog.com) (geminicli.com) In plain terms, the main agent becomes a coordinator: one helper checks tests, another studies a codebase, another drafts documentation, and the top-level agent pulls the pieces back together. Google’s subagent docs describe separate system instructions, tool access, and context for each helper. (developers.googleblog.com) (geminicli.com) That structure addresses a basic problem in coding assistants: a single long-running chat gets crowded with every file, tool call, and intermediate step. Google Cloud wrote in late 2025 that the choice between subagents and “agents as tools” affects overhead, scaling, and efficiency because the wrong pattern can force systems to pass around too much history. (cloud.google.com) The shift has been building for months. Google expanded Gemini Code Assist’s “Agent Mode” in 2025, and OpenAI had already published an “Orchestrating Agents” cookbook that framed handoffs as one agent transferring an active conversation to another specialist. (developers.googleblog.com) (developers.openai.com) This week’s burst of attention came from product updates and creator coverage arriving together. A YouTube video published on April 16 highlighted Gemini subagents and an OpenAI Agents Software Development Kit update with sandboxing and a new harness in the same roundup. (youtube.com) (developers.openai.com) OpenAI’s current docs also add a more concrete reason to break work apart: sandbox agents in the Python Software Development Kit can run in container-based environments with files, commands, packages, ports, snapshots, and memory. That makes the specialist-worker model easier to apply to coding jobs that need isolated execution. (developers.openai.com) Google is making a similar argument from the command line. Its Gemini Command Line Interface docs say subagents are suited to “deep codebase analysis, documentation lookup, or domain-specific reasoning” without cluttering the main agent’s toolset. (geminicli.com) (github.com) The result is a quieter change than a new model launch, but a visible one in developer tooling: the assistant is becoming the foreman, and the work is moving to a small crew behind it. (developers.openai.com) (developers.googleblog.com)