Everything it does

Find what you want to do, and go straight to the page that does it.

ARun it your way

Ready-made loops Any coding agent Model and effort Two accounts of one CLI Fall back Skills From a script In CI From Python

BWhile it runs

Talk into a turn Side questions Watch every agent Answer, or step away What it costs A budget on every run Stop it Leave it running

CWhere the work lands

In a container On another machine What an agent may touch

DAfter a run

Pick it up One timeline Hand it to somebody Crash reports

EWriting a flow

A loop in plain Python Params of its own Many conversations at once Answers as typed data The agent decides it is done React to each moment Tools that call the flow Ask the person Cap a single turn Call another flow Branch a conversation Worktrees and copies Test without a model Publish it

Run it your way

You can	Go to
Ready-made loops. A Ralph loop, a reviewer loop, three lanes at once.	Flows
Any coding agent. Claude Code, Codex, Cursor, Kimi and more, most on the login they have. Add any Agent Client Protocol CLI.	Accounts Every coding agent you have
Model and effort. Each agent's model, and how hard it thinks.	Efforts
Two accounts of one CLI. A subscription and a gateway side by side, each with its own login.	Accounts Two accounts of one CLI
Fall back. A failed turn is tried again, then handed along a chain of other accounts, CLIs or models.	Falling back Two accounts of one CLI
Skills. The skills each agent loads. A flow can bring its own.	Skills
In a browser. The run going, every run before it, starting one and what they spent, on a page.	In a browser Prompt, browser, script or Python
From a script. A flow from a shell script or a cron job, no interface.	Run it unattended Prompt, browser, script or Python
In CI. A flow on a schedule that opens a pull request.	Humanize in CI
From Python. Humanize driven by a program of your own.	SDK reference Prompt, browser, script or Python

While it runs

You can	Go to
Talk into a turn. Correct an agent mid-turn without stopping it.	Talking to a running turn Talk into a running turn
Side questions. Ask a running flow what it is up to, without interrupting.	Side questions
Watch every agent. Who is working, for how long, and who handed over to whom.	Watching a run
Answer, or step away. Answer when asked, or say you are away so nothing waits.	Questions, Being away When a flow asks you
What it costs. Tokens, money and rate for every agent, live.	Cost and rate
A budget on every run. Time, cost or output tokens; the first limit reached stops it.	Run it unattended A budget on every run
Stop it. From the keyboard, and by force if it will not stop.	Stopping
Leave it running. Close the terminal; open Humanize in the same directory to find it again.	Daemon reference Close the terminal, keep the run

Where the work lands

You can	Go to
In a container. A toolchain you have not got, with your project at the same path.	Containers
On another machine. Commands run on the build box; the agent and its login stay here.	Remote execution Work on another machine
What an agent may touch. What a flow lets each agent do. Nothing is put to you.	Permissions, Security

After a run

You can	Go to
Pick it up. Carry a stopped run on from where it stood.	Picking a run up Pick up where it stopped
One timeline. Every agent and sub-agent on one clock in Perfetto, programs too if profiled.	Tracing Every agent on one timeline
Hand it to somebody. A whole run packed into one archive.	Exporting a run
Crash reports. Send them and feedback, or never. You are asked once.	Reporting

Writing a flow

Python, for the weaver: whoever writes the flow.

You can	Go to
A loop in plain Python. An async function, with its agents declared by role.	Writing a flow A loop in plain Python
Params of its own. Typed settings the prompt and the command line both fill in.	Params of its own
Many conversations at once. Fan out as wide as the work needs.	Many turns at once Many conversations at once
Answers as typed data. A pydantic model back, not a paragraph.	Answers in a shape Answers as typed data
The agent decides it is done. A goal it keeps at until it judges it met.	Goals The agent decides it is done
React to each moment. Your code before a tool, on a prompt, as a turn stops.	Hooks React to each moment of a turn
Tools that call the flow. The agent calls a tool of the flow's, runs a command line of the flow's from its shell, or asks it, mid-turn; the flow's code answers.	The agent asking the flow
Ask the person. A question to whoever is at the prompt, as one of the flow's agents.	The person as an agent When a flow asks you
Cap a single turn. Stop one turn at a time, money or token limit.	Flows reference Cap a single turn
Call another flow. Another flow as a step, under what is left of your budget.	A flow that calls a flow
Branch a conversation. Fork it and try more than one way on.	Branching a conversation
Worktrees and copies. A worktree per task, a throwaway copy, or a scratch directory.	Worktrees, copies and scratch
Test without a model. Scripted agents: milliseconds a test, nothing spent.	Testing a flow
Publish it. A flow in a repository of its own, listed in a flowverse anybody can install it from.	Flowverses

Which CLI does what

Most of the map works the same on every CLI. Three things do not:

CLI	Talk into a turn	Goals	Trace
Claude Code	✓	✓	✓
Codex	✓	✓	✓
Kimi Code	✓	✓	✓
pi, Oh My Pi	✓		✓
DeepSeek Harness		✓	✓
Antigravity, Grok Build, MiMo Code, MiniMax Code, opencode, Qwen Code			✓
Cursor Agent
a CLI you added yourself

Elsewhere, what you type mid-turn is taken as the next turn. Everything a flow can ask of each CLI is in the agents reference.

User GuideRunning flows: settings, watching, stopping, picking up.Writing FlowsWriting flows of your own, in Python.ReferenceEvery command, setting and flow API, in full.