Efforts

An agent is a backend, a model and an effort — how hard to think.

claude / claude-opus-4-8 : high
  │           │            └── effort
  │           └── model
  └── backend

The word is the backend's own, not humanize's, so the values differ. humanize does not check an effort against a list: a value your account has and this page does not still works.

Backend	Efforts, hardest first
Claude Code	ultracode, max, xhigh, high, medium, low
Codex	ultra, max, xhigh, high, medium, low — each model takes its own subset
Kimi Code	max, high, medium, low, each also as swarm…
pi	max, xhigh, high, medium, low, minimal, off
opencode, mimocode	the model variant: xhigh, high, medium, low, minimal

ultracode is Claude Code's xhigh thinking with the turn opted into orchestrating a fleet of its own. It is more work than any single-agent effort, which is why it sits above max. It is real and undocumented, so no listing the CLI answers with will ever name it — humanize keeps it anyway.

Kimi Code's effort says how wide as well as how hard. max is one agent; swarmmax is the same thinking at the width of a fleet of subagents. The prefix is exported as hmz.agents.SWARM.

pi's off is the model asked not to think at all. That is an effort like any other here: the least of them, not the absence of a setting.

Codex's models differ from each other — gpt-5.6-sol takes ultra, gpt-5.5 does not — so the interface offers each model only the efforts it takes.

Setting it

command line

hmz exec -f ralph_loop -a claude/claude-opus-4-8:high "fix the build"
hmz exec -f ralph_loop -a cli=kimi,model=kimi-code/k3,effort=swarmmax "fix the build"

Python

ClaudeCodeAgentConfig(model="claude-opus-4-8", effort="high")

At the prompt it is on the model step of /agents: ←/→ adjusts the effort, and ctrl+w turns swarm mode on for a model that has one.

Moving it while the flow runs

A config is frozen — a session resumes under the settings it opened with, and a config that changed mid-flow would silently split one conversation across two models. The effort is the one setting a flow may move as it goes:

agents.builder.effort = "low"       # every session of this agent, from its next turn
session.effort = "max"              # this conversation alone
session.effort = ""                 # and back to whatever the agent runs at

Reading it back is the same property:

agent.config.effort	what the agent was configured with
agent.effort	what its turns actually run at

It takes hold on the next turn. The turn already under way keeps the effort it started at: a model does not think harder halfway through an answer.

Each backend carries it its own way. Codex, Kimi Code, opencode and mimocode take the effort with each turn, so the next turn simply carries the new one. Claude Code takes it as an argument of the process it is held open as, so moving it ends that process and resumes the conversation in one started at the new effort. pi has a command for it, and is told.

A swarm prefix moves with it: agent.effort = "swarmmax".

What to steer by

The reading that responds to effort is juice() — output tokens an average turn of the model came out with. A model asked to think harder writes more in each answer and takes longer over it, so that average is what an effort moves.

agent.juice(over=60)

It is what official/fixed_juice_ralph governs on: a Ralph loop that moves the effort a rung a round to hold the agent to a target. See Cost and rate.

See also

• Cost and rate
• Permissions — the other thing set on the model sheet
• Agents › Efforts