Glossary · term80 of 202
fan-out
The pattern where one delegated task spawns multiple concurrent subagents, and each of those can spawn more of its own. Depth is usually capped, but breadth often is not, so total agents alive can multiply far past what anyone intended.
also written as fan out
Related terms
drained planA usage plan whose budget has been exhausted mid-run, surfacing to a nested fan-out of subagents as repeated API errors rather than a clean stop.view termexecution subagentA subagent given a narrow task with a fixed output shape and a written boundary, as opposed to open-ended reasoning or planning work. Its job is to follow the spec, not to expand scope.view termmixed-model benchmarkA test comparing combinations of different models working together against a single model working alone on the same task, used to judge whether splitting work across models pays off.view termno-op turnA turn where a parent agent produces no new work because it is only waiting on background subagents to finish, yet still consumes a turn of the conversation.view termorphaned agentA spawned subagent whose parent task record is gone, so it keeps running and consuming resources while showing as untrackable and unable to be stopped.view termsubagent bootstrapThe startup cost of spinning up a subagent, including loading its own system prompt and full set of tool schemas, repeated on every turn it takes.view termtask toolThe mechanism an agent uses to spawn a subagent to handle a piece of work, the same primitive that lets subagents recursively spawn further subagents.view term
At a glance
- cited by
- 4 posts
- categories
- 3
- first used
- may 2026
30 tips and tricks with Claude Code CLI on WSLClaude Code on WSL returns fewer search matches than exist when a repo sits on /mnt/c. A 2,000-file benchmark, plus 30 fixes for the WSL boundary.
Why expensive models make terrible execution subagentsSubagent model selection has one costly default. Reasoning models obey their own traces 15.6% of the time, so here is where to pin model and effort.
Token-Compression Skills Do Not Survive MeasurementToken compression skills report savings the invoice never shows. Cache reads carry 87 percent of an agent bill, and cutting tokens can push the cost up.
Cut MCP Round-Trip Overhead by Looping Inside the ToolMCP round-trip overhead can balloon a task from 2,000 to 150,000 tokens. Move the loop inside one MCP tool to collapse N calls into a single round-trip.