Effort level
A control on an agentic coding harness that sets how much work the agent does on a request — files read, tools called, steps taken, and how much it explores and verifies before deciding a task is finished. In Claude Code it is a distinct axis from model selection: the model fixes the capability ceiling, effort governs how thoroughly the agent works inside it.
Not just thinking time
The common misreading is that effort only buys more reasoning tokens before an answer. The Anthropic framing is broader — effort scales the whole task loop: exploration breadth, tool use, and the bar for calling work complete claude-code-model-effort-levels. Mechanically, the effort level is passed as an input alongside the prompt; the behaviour for each level was learned during training and is “baked into the frozen weights.” So — unlike switching models, which changes capability but not how many tokens are generated — raising effort does increase token generation, because the agent does more.
The two knobs, and which to reach for
- Too little work (skipped files, no tests run, task abandoned mid-way) → raise effort.
- Confidently wrong with clear context → a capability limit; upgrade the model, don’t just raise effort.
- Otherwise leave the default, which is tuned to what most users would want to spend and is best treated as a standing preference rather than a per-task decision.
Where it sits
Effort is the harness-level expression of the reliability discipline this spoke keeps returning to — the same instinct as phase gates that forbid skipping tests (agentic-coding-harness) and verification loops, but exposed as a single tunable instead of hand-built scaffolding. It is orthogonal to the model-tier choice tracked in claude-code-best-practices and the cross-wiki model market (claude-opus-4-8, claude-fable-5).