Scheduled Jobs

Cron jobs publish an InboundMessage to the same bus as user messages — they flow through the identical middleware pipeline and produce channel output.

Cron is off by default. Enabling it with the default SQLite data store also requires sqlalchemy and aiosqlite:

LANGCLAW__CRON__ENABLED=true
uv add sqlalchemy aiosqlite   # needed for the default sqlite cron data store

Schedule via the agent

Once cron is enabled, the agent gets a cron tool. Users schedule jobs in plain language:

Schedule a daily HN digest every morning at 8am
Schedule the landscape workflow for "agent frameworks" every Monday

The agent translates this into a cron job via the cron tool.

Schedule a saved workflow

Saved workflows scheduled via cron run their frozen body without any LLM call — deterministic, zero LLM cost:

Schedule workflow 'digest' every day at 9am

On fire, the workflow runs verbatim. If the .js file is deleted, the job self-disarms.

Manage jobs

/cron list          → all scheduled jobs
/cron view <id>     → show job details
/cron remove <id>   → remove a job

Where output goes

A cron job has no live conversation, so it captures its delivery target at schedule time: when the agent's cron tool creates a job, it reads the current LangclawContext and stamps the channel, user_id, and chat_id into the job. On every fire the result is delivered back to that same channel/chat. Schedule a digest from your Telegram chat and it arrives in that chat.

Internals

Cron is backed by APScheduler v4. Jobs are persisted to APScheduler's own data store (LANGCLAW__CRON__DATA_STORE__BACKEND, sqlite by default — a separate store from the conversation checkpointer), so they survive restarts. Each fire also stamps agent_name into InboundMessage.metadata (from the context at schedule time), so jobs always run against the right agent even if the user switches later.