Subagents

The big advantage is isolation.

When a primary agent invokes a subagent, that work happens in a child session instead of cluttering the context of the parent session with every search, command, and intermediate thought.

Subagents can also be cost-efficient because the isolated session can use a different model. For example, you might want your explore subagent to use a small, cheap model for searching around the codebase, then summarize its findings and inject them back into the parent session for your expensive model to analyze.

They can also be faster because subagents can run in parallel. When it launches a subagent in the background, your primary agent can keep working and receive the result when it's ready. Foreground calls wait for the subagent to finish.

Subagents are mostly called autonomously by your primary agent, just like any other tool. However, you can also invoke them manually with the @ symbol:

@explore go find all the files related to authentication and summarize their contents for me

Have a subagent find a gap in Sloperator's lexer tests.

1. Tell the primary Build agent to delegate a search to explore for a lexical error without regression coverage. Inspect the child session's work and its summary. (You can click the subagent's name in the session to see its work. It's collapsed by default.)
2. Tell the primary Build agent to add a test for the uncovered error in tests/test_lexer_regression.py.

Tell OpenCode to run and submit the CLI tests from the project root.