Efforts

An agent is a CLI, a model and an effort: how hard the model thinks before it answers. Use this page to choose one, when a task needs more thought than you are getting or a loop is spending more than the work is worth. By the end you will set an effort on the command line and at the prompt, and know which words each backend takes.

Try it

The effort is the word after the last colon of an agent:

hmz exec -f chat -a assistant=claude/claude-haiku-4-5-20251001:low \
    "Name the three primary colours, one line."

● Red, yellow, and blue are the three traditional primary colors.
✻ input 10 · output 136 · cache_read 13.9k · cache_write 6.9k · $0.01 · claude-haiku-4-5-20251001 · assistant

Before you start

• Humanize is installed, with at least one coding agent CLI signed in.
• You know which CLI and model you want. The effort is the last thing you choose, because which efforts there are depends on both.

What an effort is

Each coding agent CLI has its own words for how long a model reasons: Claude Code says ultracode to low, Codex says ultra to low, pi goes down to off. Humanize takes the CLI's own word and passes it on, so an effort always means what that CLI means by it. It does not translate one backend's high into another's.

• A harder effort reasons longer. It writes more output tokens, so it costs more and takes longer, and it is what a hard bug or a large refactor wants.
• An easier effort answers sooner and cheaper. It suits a loop that repeats a mechanical step many times.
• auto says nothing about effort at all, and the model runs at the CLI's own default.

An agent keeps the effort you gave it for the whole run. A flow can read the effort, but it cannot change it, so a flow never makes an agent think harder than you asked.

Example: pick an effort, and have a wrong one refused

You want Claude Code's cheapest model to answer quickly. First, try a word Claude Code does not have:

hmz exec -f chat -a assistant=claude/claude-haiku-4-5-20251001:turbo "hi"   # ①
echo $?                                                                    # ②

hmz exec: error: assistant=claude/claude-haiku-4-5-20251001:turbo: claude cannot be asked to think at 'turbo'; expected one of ultracode, max, xhigh, high, medium, low
2

Then use one from the list it gave you:

hmz exec -f chat \
    -a assistant=claude/claude-haiku-4-5-20251001:low \
    "Name the three primary colours, one line."                            # ③

● assistant is working
● Red, yellow, and blue are the three traditional primary colors.
✻ input 10 · output 136 · cache_read 13.9k · cache_write 6.9k · $0.01 · claude-haiku-4-5-20251001 · assistant   ④
Red, yellow, and blue are the three traditional primary colors.
✻ Worked for 3s · assistant

What each part means

1. :turbo is read as the effort because it comes after the last colon. The CLI is read up to the first /, so a model with slashes or colons in its name needs no quoting.
2. Exit status 2 means the line was refused before anything ran: no agent started, and nothing was spent. The message names the CLI and lists its whole ladder, hardest first.
3. :low is the bottom of Claude Code's ladder. For a one-line answer, it is plenty.
4. The ✻ line is what the turn cost. output 136 is the tokens the model wrote, reasoning included. Run the same line at :high and that figure is the one that grows. See Cost and rate.

Example: set it at the prompt

In hmz, open /flow, choose the flow, then the agent's role. The agent sheet has one row per choice, and the effort is the last of them:

  hmz › /flow › Installed › chat › Set up assistant
  Configure this agent: select its CLI, account, model, and reasoning effort.

  ╭──────────────────────────────────────────────────────────────────────────╮
  │ cli                                                             claude ▸ │
  │   coding agent CLI to use                                                │
  │──────────────────────────────────────────────────────────────────────────│
  │ account                                                       as local ▸ │  ①
  │   account to run as                                                      │
  │──────────────────────────────────────────────────────────────────────────│
  │ model                                        claude-haiku-4-5-20251001 ▸ │  ②
  │   model to use                                                           │
  │──────────────────────────────────────────────────────────────────────────│
  │ effort                                                             low ▾ │  ③
  │   reasoning effort                             ╭─ effort ─────────────╮  │
  │                                                │ ultracode            │  │
  │                                                │ max                  │  │
  │                                                │ xhigh                │  │
  │                                                │ high                 │  │  ④
  │                                                │ medium               │  │
  │                                                │ low ✔                │  │
  │                                                ╰──────────────────────╯  │
  ╰──────────────────────────────────────────────────────────────────────────╯

                                                                          Save

  enter choose   esc close

What each part means

1. account is who the turns run as. as local is the CLI as this machine is signed in. See Accounts.
2. model lists, for each model, the efforts it takes, so you can see a model's ladder before you pick it: claude-haiku-4-5-20251001 ultracode, max, xhigh, high, medium, low.
3. ▾ marks a row whose values drop under it. enter or a click shows the model's whole ladder, hardest first, with ✔ on the effort in force.
4. The list: ↑ ↓ and enter, or a click, pick an effort; esc or a click off the list keeps the one you had.

Press Save under the sheet (tab, enter), then save on the flow. The line above the prompt then reads assistant · claude/claude-haiku-4-5-20251001:low.

Check that it worked

• Under hmz exec, a line that runs at all was accepted: a refused effort exits 2 before anything starts.
• At the prompt, the agent line above the editor ends in the effort, as in …:low.
• Compare the output figure on the ✻ line of two runs at different efforts. The harder one writes more.

Every backend's ladder

Each CLI has its own words, so the ladder depends on the backend. Pick one, then click a rung or type any word to see what -a makes of it:

Claude Codehardest first · click a rung

or type a word

$ hmz exec … -a agent=claude/MODEL:high

✓ takentaken

Claude Code never lists ultracode itself; Humanize offers it anyway.

A simulation of the check hmz exec -a makes, drawn from Humanize’s own table of backends. Nothing here runs a CLI.

A word that is not on the backend's ladder is refused before anything runs. On Codex, Cursor Agent, MiniMax Code and Oh My Pi each model takes only some of the rungs, and the agent sheet offers each model its own.

The same ladders, as a table

Backend	Efforts, hardest first
claude	ultracode not listed by Claude Code, max, xhigh, high, medium, low
codex	ultra, max, xhigh, high, medium, low subset per model
cursor-agent	max, xhigh, extra-high, high, medium, low, minimal, none subset per model
dsh	max, high, low, off
grok	xhigh, high, medium, low
kimi	max, high, medium, low each also as swarm…
mcode	max, xhigh, high, medium, low only MiniMax-M3.1-Flash-Preview
mimo, opencode	xhigh, high, medium, low, minimal: the model's variant
pi	max, xhigh, high, medium, low, minimal, off
omp	max, xhigh, high, medium, low, minimal, off subset per model
qwen	max, xhigh, high, medium, low, none
agy	high, medium, low
a CLI added on the Accounts page of /settings	any word not sent
every backend	auto

Variations

auto: no effort at all

auto tells the CLI nothing about how hard to think, so the model runs at its own default. Every backend takes it. Use it for a model that has no rungs, such as Cursor's composer-2.5 or many models behind a gateway:

hmz exec -f ralph_loop -p budget.cost=5 \
    -a agent=cursor-agent/composer-2.5:auto "fix the build"

auto is not the bottom rung. The off of pi, Oh My Pi and DeepSeek Harness, and Qwen Code's none, ask the model not to think at all; auto asks nothing. On the agent sheet, a model with no rungs shows no effort to pick, and runs as auto.

Kimi's swarm mode

On Kimi Code, the effort also says how wide a turn runs. swarm in front of a rung runs the same thinking as a fleet of subagents, so swarmmax is max, run wide:

hmz exec -f ralph_loop -p budget.cost=5 \
    -a agent=kimi/kimi-code/k3:swarmmax "fix the build"

On the agent sheet it is a row of its own, swarm, shown for a model that has it.

Two roles, two efforts

Each role has its own agent, so each has its own effort. A cheap maker and a careful checker is a common split:

hmz exec -f rlar -p budget.cost=10 \
    -a actor=claude/claude-opus-5:medium \
    -a reviewer=codex/gpt-5.6-sol:xhigh "fix the build"

Pitfalls

• A new model starts at its hardest rung. Choosing another model on the sheet keeps the effort if that model takes it, and otherwise starts at the hardest one it does, such as ultracode on Claude Code. Step it down before you save if you did not mean that.
• cannot be asked to think at '…'. The word is not on that CLI's ladder. Pick one from the list in the message, or auto. See Troubleshooting.
• cursor-agent lists no … this account runs … as …. Cursor writes the effort into the model's id, and your account does not offer that model at that effort. Pick an effort the message lists.
• The same word, different cost. high on one backend is not high on another. Compare the output figure on the ✻ line, not the word.

Next steps

• Cost and rate: what a harder effort costs
• Accounts: the account a model runs as
• Run it unattended: the rest of an -a
• Agents › Efforts: the whole reference