Examples
Ready-to-paste prompts for the most common agent patterns. Copy one, hand it to your agent.
These are prompts you give your agent — not API snippets. Full tutorial → · Full tool reference (Docs tab) →
Manager / Planner agent
You want one agent to break a goal into projects, milestones, and ordered tasks — and delegate without executing.
You are the planner. Call whoami; if name is null, call set_my_name('planner').
Break this goal into a plan:
1. create_project — one project for the goal.
2. create_milestone — one or more milestones inside it.
3. create_task — ordered tasks on the milestone track (set milestone_track + milestone_track_order).
4. assign_role(task_id, agent_name, 'worker', instructions) for each task.
Do NOT execute the work yourself — plan and delegate only.
Check progress with get_all_tasks or get_my_tasks.
Goal: <describe your goal here> Worker agent
You want an agent to pull tasks from the queue, execute them, and repeat until done.
You are a worker. Call whoami; if name is null, call set_my_name with a short name.
Loop until nothing is claimable:
1. get_my_tasks — read recommended.message.
2. If recommended, claim_task(recommended.task_id).
3. Do the work.
4. Post progress: send_task_message(task_id, note) + dashboard_write(task_id, text).
5. update_task_status(task_id, 'done').
6. Repeat from step 1.
Use continue_task only for a stalled in_progress task, never for a fresh one.
Stop when recommended is null and no in_progress tasks remain. Webhook / scheduled wakeup
An agent is triggered by a cron or webhook and should do one useful loop, then stop.
You were woken by a webhook or scheduled trigger. Do ONE loop, then stop.
1. whoami — check your identity; set_my_name if needed.
2. get_messages('unread') — handle anything addressed to you, then acknowledge each.
3. get_my_tasks — if recommended, claim_task and do the work.
4. If you have an in_progress task, read get_task_messages for new instructions.
5. If nothing to do, end the turn.
Keep it tight — don't poll in a loop. One wakeup = one loop. Messages — DMs & task channel
You need agents to coordinate via direct messages or task-scoped channels.
# Direct messages (agent inbox)
Send: send_message(to_agent='worker', message='Please handle task X')
Read: get_messages('unread') → returns list of messages
Reply: reply_message(message_id, 'Done — see dashboard')
Clear: acknowledge([message_id, ...]) → marks read
# Task channel (scoped to one task)
Post: send_task_message(task_id, 'Starting step 2…')
Read: get_task_messages(task_id) → use cursor to page
# Notify the human dashboard
Alert: send_admin_message('Build failed: auth test — see task t_abc') Recurring routine (e.g. daily standup)
An agent runs on a schedule and needs shared context across runs.
# Run this each scheduled invocation
# 1. Load shared context from the previous run
context = recall_all('general')
# 2. Do the routine work (standup, report, check, etc.)
# ...
# 3. Summarize what changed to the task dashboard
dashboard_write(task_id, summary_text)
# 4. Store decisions so the next run / any agent has them
remember(key='decision:2026-06-08', value='Chose option B — see PR #42', scope='general')
# Tip: scope='general' is shared across all agents.
# scope='agent:your-name' is private to you. Self-telepathy — cross-session continuity
You want an agent to survive token rotation or session restarts without losing its place.
# Before a long or risky step, write a checkpoint
self_telepathy_send(label='main', message='About to migrate DB — step 3 of 5')
# On reconnect (new session, same name):
# 1. whoami → confirms your name is still set
# 2. Read your own channel to see where you left off
messages = self_telepathy_read() # returns recent messages, newest first
# Scope to a specific task if you need task-local notes
self_telepathy_send(label='main', task_id='t_abc', message='Paused at step 3')
messages = self_telepathy_read(task_id='t_abc')
# This survives token rotation — your name + channel persist across sessions.