Autonomous Hands

Hands are autonomous workflows that the agent executes on a schedule, with their own budget, tools, and approval gates. Think of them as background jobs that the agent runs independently — checking in with you only when approval is needed.

What is a Hand

A Hand is a named, scheduled task with:

A prompt — what the agent should accomplish
A schedule — cron expression for when it runs
Tools — which MCP tools it can use (scoped, not all tools)
A budget cap — maximum spend per run
An approval gate — whether to pause for human approval before acting

hands:
  - name: "linkedin-prospector"
    schedule: "0 9 * * 1-5"          # Weekdays at 9am
    prompt: "Find LinkedIn prospects matching Cronoss ICP and add to CRM"
    tools: ["browser_navigate", "browser_extract", "clickup"]
    budget_cap_usd: 0.50
    output_channel: "default"
    approval_required: true
    approval_target: "Esteban"

Lifecycle

Each Hand run follows a state machine:

pending → planning → awaiting_approval → executing → completed
                                                   → failed
                         → rejected

Pending — the cron fires, the Hand is queued
Planning — the agent creates a numbered plan with steps
Awaiting Approval — if approval_required: true, the plan is sent to the approval target for review
Executing — approved plan steps execute sequentially
Completed/Failed — final state, results sent to output_channel

Plans

Before executing, a Hand creates a plan — a numbered list of steps the agent intends to take. Plans provide transparency and control.

Plan: linkedin-prospector (Run #42)
1. Search LinkedIn for CTOs at EdTech companies (50-200 employees)
2. Extract top 10 profiles with contact info
3. Cross-reference with existing CRM contacts
4. Add 5 new prospects to ClickUp "Leads" list
5. Draft personalized outreach messages

Budget estimate: $0.35 / $0.50 cap
Approve? [yes/no]

Plans support comments — you can add notes before approving, and the agent adjusts its execution accordingly.

Shared Plans

Plans can be shared with team members for collaborative review:

Share via link (token-based access)
Recipients can comment but only authorized roles can approve
Audit trail tracks who approved what

Budget Control

Each Hand has an independent budget cap. The agent tracks spend per run and stops if the cap is reached:

hands:
  - name: "weekly-report"
    budget_cap_usd: 0.20
    # ...

Budget enforcement happens at the tool execution layer. If a tool call would exceed the remaining budget, it’s skipped and the Hand reports what it couldn’t complete.

Defining Hands

Via Config

Add entries to the hands array in config.yaml:

hands:
  - name: "daily-standup"
    schedule: "0 9 * * 1-5"
    prompt: "Summarize yesterday's git commits and today's ClickUp tasks"
    tools: ["github", "clickup"]
    budget_cap_usd: 0.10
    output_channel: "default"
    approval_required: false

Via Chat

Tell the agent to create a Hand:

> Create a Hand called "weekly-report" that runs every Friday at 5pm,
  summarizes the week's PRs and tasks, and sends it to WhatsApp.
  Budget: $0.30. No approval needed.

The agent updates config.yaml and registers the new Hand automatically.

Via Portal

The web portal provides a CRUD interface for Hands with fields for name, schedule, prompt, tools, budget, and approval settings.

Hand Templates

Pre-built Hand definitions that can be installed with one command:

> Install the "weekly-report" template

Templates provide tested prompts, tool configurations, and sensible defaults for common workflows. They’re stored in internal/tool/templates/ and installable via the built-in template_install tool.

Monitoring

Hand runs are logged in a SQLite table (hand_runs) with:

Run ID, Hand name, start/end timestamps
Status (completed, failed, rejected)
Token usage and cost
Plan steps and their outcomes
Approval history

View run history via the web portal or by asking the agent: “Show me the last 5 runs of the linkedin-prospector Hand.”