human

human is a two-phase pattern with a pause for human input between phases. Phase 1 prepares a draft or question, pauses for the human to respond, then phase 2 continues with the human’s input.

When to use it

• Approval workflows (draft → human approves → finalize)
• Interactive research (agent gathers context → human refines direction → agent completes)
• Tasks that need a human checkpoint before committing to a direction
• Any workflow where human judgment should be incorporated mid-run

File structure

agents/my-human/
├── AGENT.md    ← pattern: human, config
└── resume.md   ← system prompt for phase 2 (after human input)

AGENT.md

---
name: draft-approver
description: Drafts a document and requests human approval before finalizing.
version: "1.0.0"
pattern: human
call:
  model:
    role: writer
  uses:
    - search-web
---

You are a document writer. Given the requirements:
1. Research the topic thoroughly
2. Write a complete draft
3. Present the draft to the human for review

Call finish with a channel ID as the key and the draft as the value.
The channel ID becomes the identifier for the human pause channel.

Example: finish(key="draft-review", value="Here is the draft:\n\n...")

resume.md

You receive the human reviewer's feedback on your draft.

Incorporate all requested changes. If the reviewer approved without changes, return the original draft.
If they requested edits, make precisely those edits.

Call finish with key="done" and the final document as value.

Scaffold

tama add human my-agent

How it works

1. Phase 1 — AGENT.md body runs as a react loop; when ready, calls finish(key="channel-id", value=<draft>)

2. Pause — the runtime opens a channel identified by the key; in development, waits for stdin; in production, exposes an HTTP endpoint for your app to POST the response

3. Phase 2 — resume.md receives the human’s response via start(); runs as a react loop; calls finish(key="done", value=<final result>)

The channel key

The key in phase 1’s finish is a channel ID — not a routing word. It:

• Identifies which channel to open for human input
• Appears in the trace so you can correlate human responses to runs
• Enables parallel pauses: different unique IDs can pause and resume independently

Development vs production

Development (tama run CLI): the runtime pauses and prompts stdin for input.

Production (Docker): the runtime pauses and exposes a channel ID. Your application POSTs the human response to the runtime’s HTTP API when available.