{"schema_version":"newruntime-agent-readable-v0.2","type":"capability_delta","stable_id":"capability_delta:claude-code-adds-independent-effort-and-context-controls-for-subagents","slug":"claude-code-adds-independent-effort-and-context-controls-for-subagents","title":"Claude Code adds independent effort and context controls for subagents","description":"Builders can now specify a subagent's effort level and configure earlier context compaction. Workflow agents can also use a shared model override without changing other subagents.","retrieval_nugget":"As of 2026-11-17: The configuration controls described in October remain supported by the evidence provided; no newer Claude Code release is listed. An archive item gives a concrete use of subagent-specific auto-compaction, but does not establish a later system change. Builders can now specify a subagent's effort level and configure earlier context compaction.","status":"published","observed_at":"2026-10-09","record_date":"2026-10-09","date_kind":"observed_at","before_capability":"The Agent tool lacked an explicit effort parameter, and subagent definitions lacked a separate auto-compaction window.","enabling_change":"Version 2.1.292 adds effort to the Agent tool; 2.1.296 adds autoCompactWindow to subagent definitions and a workflow-only model override.","after_capability":"Builders can request subagent effort levels, compact subagent context earlier than the main conversation, and override workflow-agent models independently of other subagents.","human_consequence":"Builders gain explicit controls for tuning delegated work rather than applying the same execution settings across the session.","verification_level":"source-inspected","buildability_class":"B4","bottleneck":"The entries provide configuration controls but no measurements of their effects on task quality, latency, or cost.","related_shift":"goal-scoped-agent-loops","related_posts":[],"topics":[],"source_urls":["https://github.com/anthropics/claude-code/releases/tag/v2.1.292","https://github.com/anthropics/claude-code/releases/tag/v2.1.296"],"featured":false,"supporting_evidence":[{"title":"claude-code 2.1.292 Claude Code 2.1.292","url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.292","kind":"external","observed_at":"2026-10-06"},{"title":"claude-code 2.1.296 Claude Code 2.1.296","url":"https://github.com/anthropics/claude-code/releases/tag/v2.1.296","kind":"external","observed_at":"2026-10-09"}],"routes":{"html":"https://newruntime.com/deltas/claude-code-adds-independent-effort-and-context-controls-for-subagents/","markdown":"https://newruntime.com/deltas/claude-code-adds-independent-effort-and-context-controls-for-subagents.md","json":"https://newruntime.com/deltas/claude-code-adds-independent-effort-and-context-controls-for-subagents.json"},"source_format":"knowledge-ledger","published_at":"2026-11-17T05:08:11.198Z","auto_created":true,"editorial_provenance":{"desk":"knowledge-revision","provider":"codex","model":"gpt-6.1-sol","created_at":"2026-11-17T05:08:11.198Z","week_start":"2026-10-05","review":"auto-created from reviewed evidence; the owner may revise or retire it"},"current_state":{"as_of":"2026-11-17","status":"holding","summary":"The configuration controls described in October remain supported by the evidence provided; no newer Claude Code release is listed. An archive item gives a concrete use of subagent-specific auto-compaction, but does not establish a later system change.","text":"Claude Code builders can set effort for subagents, configure a separate auto-compaction window, and apply a model override to workflow agents without changing other subagents. The cited releases are 2.1.292 and 2.1.296; the items provide no newer release information.\n\nAn archived example describes setting a reviewer subagent’s autoCompactWindow between 100,000 and 1,000,000 tokens while leaving the main conversation’s window unchanged. The evidence does not report effects on quality, latency, or cost, so builders have controls but no demonstrated performance tradeoffs here.","changed_since":[{"text":"An archived example describes using autoCompactWindow to set an individual reviewer subagent’s compaction window while leaving the main conversation unchanged.","sources":["https://dev.to/gabrielanhaia/a-read-only-reviewer-subagent-in-claude-code-3-tools-haiku-55-a-100k-compact-window-95j"]}],"outdated_in_original":[],"open_questions":["Do the effort and model controls measurably affect task quality, latency, or cost?","Have newer Claude Code releases changed or expanded these controls?"],"watch_next":"A newer Claude Code release documenting changes to subagent effort, compaction, or workflow-agent model overrides would change this assessment.","sources":[{"title":"A Read-Only Reviewer Subagent in Claude Code: 3 Tools, Haiku 5.5, a 100K Compact Window - DEV Community","url":"https://dev.to/gabrielanhaia/a-read-only-reviewer-subagent-in-claude-code-3-tools-haiku-55-a-100k-compact-window-95j","kind":"archive","observed_at":null}],"provenance":{"desk":"living-knowledge","provider":"codex","model":"gpt-6-luna","note":"Rewritten automatically; newer than the original text where they disagree."}}}
