Anthropic's "sub-agents" feature in its Claude Code coding tool, which runs multiple AI assistants in parallel, is spreading among developers as a way to break the bottleneck of sequential prompting.
Claude Code · Anthropic
Sub-Agents: Parallel AI Instances That Keep Context Clean
Each sub-agent runs as a dedicated AI instance with its own context window. Intermediate work stays isolated — only the final output returns to the parent — so specialized tasks run in parallel lanes without bloating the main session.
154+
ready-to-use custom sub-agent examples published by the community
4
agent types: Explore, Plan, general-purpose, Custom
1
separate context window per agent — parent stays clean
One lane vs. many lanes
Conventional coding repeats prompt → wait → review → re-prompt in a single window. Sub-agents run multiple lanes at once.
N
parallel specialized lanes
How delegation flows
Parent session"Use sub-agents"
→
Sub-agents runisolated context each
→
Final output onlyreturned to parent
Explore
Read-only
Exploring and searching a codebase
Plan
Read-only
Research and planning in Plan Mode
general-purpose
Explore + edit
General tasks involving edits
Custom
Configurable
Review, audit, design & more
What developers value
Parallelism boosts productivity
Prevents context-window bloat
Cheaper models curb token use
Great for refactor, debug, docs
Where it's limited
Agents can't talk to each other
Complex tasks need parent orchestration
Used sparingly unless told explicitly
Gains clearest in longer sessions
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…