Skip to content

Agents, models & effort

Each stage is run by an agent, and each agent is assigned a Claude model and a reasoning-effort level.

worca doesn’t impose fixed model tiers. The default model for each stage generally tracks the stage’s complexity — the more reasoning-heavy stages (Planner, Plan Reviewer, Coordinator, Reviewer, Guardian, and Learner) are typically set to Opus, while the build-and-test stages run a faster model. Per-agent model and max-turns are set in the Pipeline Templates editor’s Agents tab; every run picks up an active template, so the template owns this config. See Configuration precedence for the full strip-and-merge rules.

Agents run at an effort levellow, medium, high, xhigh, or max — that controls how much reasoning budget they spend. Effort is governed pipeline-wide by a mode:

ModeBehavior
adaptive (default)The Coordinator labels each task with a complexity level; the Implementer starts there and escalates on retries.
reactiveAgents start at their configured level and escalate when a loop sends work back.
disabledEach agent stays pinned to its configured level — no automatic escalation.

A ceiling (auto_cap, default xhigh) bounds how high runtime escalation can go, and the level an agent actually used is shown as a badge on each iteration in the dashboard.