concepts
Sub-agents, in isolation.
When the main thread shouldn't be polluted by a deep search or a parallel job, Anthracode spins up sub-agents. Each runs with its own context window.
When to spawn an agent
- Open-ended search — “find every place we call the deprecated logger”.
- Independent parallel work — running two refactors that don't share files.
- Heavy reading — surveying a large codebase you don't want flooding the main context.
Agent types
- explore — fast, read-only. Best for “where is X defined?”.
- general-purpose — full toolset. For multi-step research and execution.
- plan — produces a step-by-step plan; doesn't edit files.
- review — second-opinion code review of a diff.
Spawning from a session
~/anthracode session
/agent explore "where is the legacy retry policy used?"Each agent returns a single summary message. The summary describes what it intended to do, not necessarily what it did — verify before acting on the result.
Parallelism
Independent agent calls can run in parallel. The main agent waits for all of them to return, then synthesises. This is significantly faster than serial calls when the work doesn't share files.