Agent Farmbeta Join early access

District IV of XTen things

Recipes and triggers

Work that starts on time, on an event, or from one sentence.

A cup of coffee on a monthly planner with the month's goals

Some work should not wait for you to press a button: the Monday video, the folder of new footage, the report at nine. Here it starts on a clock, on an event or from one sentence, runs in turn under the limits you set, tries again when a plan runs out, and stops everywhere at once when you say so.

10 Recipes and triggers

Tasks · schedules · triggers · workflows

Work that starts by itself.

On a clock, when a file lands, or from one sentence in plain words. Switch Gatekeeper on and, before any model is woken, Jev, a model that never chats and only decides, says whether there is real work. Noise wakes nobody.

Tasks

plan what runs · see what ran — updated 03:12
Automationrunning⏸ Pause all1h8h24h🌙 Quiet off
⚙️ 0/2 running 📅 4/15 switched on 🏁 today 19✓ 0✗ 💵 $2.31/7d 🛡 2 saved ⚠️ 0 need attention
+ New task📊 Analytics
⚡ Watch trigger · New footage Trigger check
📁 Studio\inboxwatched · debounce 5 s · cooldown 60 s
🛡 Gatekeepertrigger.has-work · asks Jev · waits 3 s at most

Waiting for a firing.

🔗 Long to shortsa ready-made recipe · {file} → its video input
  1. 1Extract the soundtool · ffmpeg
  2. 2Transcriptmedia · whisper
  3. 3Pick the momentsagent · editor
  4. 4Cut the shortstool · ffmpeg, one per moment
  5. 5Final checkcheckpoint
Drop a file in the folder

Gatekeeper is off by default. Here it is on, so you can watch it say no.

IV·1

Workflows

A chain where each step hands its output to the next. A step can be an agent, an HTTP call, an image or video job, a browser run, or a stop that waits for your yes.

The Plan tab and a recipe, live

What it solves

Work that needs several agents or tools in a row means copying one result into the next by hand, and waiting at the desk to do it.

How it works · pick a step

Workflows run in VS Code. If it closes in the middle of a parallel group, Resume runs the whole group again, and an approval waits until a window is open.

Settings · 3
SettingDefaultWhat it does
workflowPrevTextChars1,500How much of the last step's text the next step is handed as {prev_text}.
agentFarm.workflowSpendCap0 (none)Once a run has spent this much, no further step starts.
agentFarm.browserApprovalaskWhether a browser step asks before it acts.
IV·1The chain1 / 5
IV·2

Schedules

Daily, weekly or once at a date and hour. A missed run is skipped, caught up, or put to you. With VS Code closed, a small service on Windows still runs them.

Schedules on the Plan tab

What it solves

A morning plan or a weekly report should start without anyone pressing a button, and a night the laptop was shut should not leave you guessing what ran.

How it works · pick a step

The service runs Claude only, and only schedules and watch triggers; hooks, webhooks and workflows wait for VS Code. Only one window runs the scheduler, so two open windows never fire a schedule twice.

IV·2Three kinds1 / 5
IV·3

Watch triggers

A file lands in a folder and an agent wakes up. A live hook, or a post to a private address on your machine, works the same way.

A watch trigger and its recipe, live

What it solves

Some work has no time of day. It starts when footage arrives, when a session ends, or when another program says so.

How it works · pick a step

The private address answers on this machine only; reaching it from outside needs a tunnel you set up yourself. A trigger that would watch its own output folder is refused, so it cannot wake itself.

Settings · 3
SettingDefaultWhat it does
agentFarm.livePort8377The local port that hooks and the private trigger addresses listen on.
agentFarm.dispatchRules[]Your own patterns that send a firing to an agent before the classifier is asked.
agentFarm.dispatchClassifieronWhen no rule matches, a cheap model picks the agent.
IV·3A file lands1 / 5
IV·4

Gatekeeper

Before a trigger wakes an agent, free local checks and then Jev, a model that never chats and only decides, say whether there is real work. Off, watching in the shadow, or on.

Gatekeeper saying no, live

What it solves

One folder watch woke opus 23 times over five nights, found nothing every time, and cost $70.90. A cooldown limits how often an agent wakes; it never asks whether there is anything to do.

How it works · pick a step

Rules come in layers, the built-in ones, yours, then a project's, the later one winning. Gatekeeper also comes as an MCP server with three tools, registered only when you ask.

Settings · 3
SettingDefaultWhat it does
agentFarm.kapici.triggeroffWhether Gatekeeper decides a trigger's firing: off, shadow or on. The same for dispatch, retry and workflow.
agentFarm.kapici.timeoutMs3000How long it may take; past it, the caller's own path decides.
agentFarm.kapici.routeautoWhich service answers for Jev, picked from the key you keep in the vault.
IV·4Free checks first1 / 5
IV·5

Say it in words

Type "run the filing batch on manager now". A cheap model turns it into a task, a schedule or a workflow, and nothing is set up until you press Create.

The words box on the Runs tab

What it solves

Setting up a job, a schedule or a chain through forms is slow, and it is the reason a good idea for an automation never gets set up.

How it works · pick a step

Its tests run a real cheap model against sentences such as "tomorrow 09:00", a two-step chain's order and picking the agent.

IV·5One sentence1 / 5
IV·6

Queue and limits

Runs wait their turn under ceilings, two at once and one per agent by default. A held run starts by itself as a slot frees.

The Automation strip, live

What it solves

Unattended jobs pile onto the same agent and the same plan. An overnight factory can eat the morning's quota before you sit down.

How it works · pick a step

One VS Code window runs the queue; the others watch. If it closes, another takes over in about 20 seconds.

Settings · 3
SettingDefaultWhat it does
agentFarm.maxConcurrentTasks2Runs at once across the fleet. 0 = no ceiling.
agentFarm.maxConcurrentPerAgent1Runs at once on one agent.
agentFarm.queueAutoAdvanceoffStart an agent's next queued task when one finishes.
IV·6Ceilings1 / 5
IV·7

A record of every run

Every run writes what it said to a file, is marked failed when it crashed, and carries its price. A job you set up often becomes a template, and finished rows move to a monthly history.

The Runs tab, live

What it solves

A run that happened while you slept is only worth something if you can read what it said, know whether it really worked and what it cost, and if the job you set up every week does not have to be typed again.

How it works · pick a step

Over the bridge, tasks.get hands an agent the archived output of a task, so one agent can read what another one's night run produced.

Settings · 1
SettingDefaultWhat it does
agentFarm.outputsMaxMB200How big the folder of task outputs may grow, in MB, before the oldest unreferenced ones go.
IV·7In the background1 / 5
IV·8

Retries and rounds

A run that fails or hits a usage limit tries again by itself, after 40 minutes or at the plan's reset, even on another account. A big job runs in rounds inside one conversation.

Runs and their retries

What it solves

A run at 03:00 that hits a usage limit should not die with nobody watching. And a job in several stages should keep one conversation, not start four.

How it works · pick a step

A retry supervisor decides how a cut-off task goes on: it resumes only when there is a session to continue, and waits only when the reset time is known.

Settings · 2
SettingDefaultWhat it does
agentFarm.taskRetry40 min, not reset-awareHow a failed run tries again: the pause, a growing back-off, the plan's reset, another account.
agentFarm.kapici.retryoffLet Gatekeeper decide how a cut-off task goes on.
IV·8A wall at 03:001 / 5
IV·9

One switch

Stop every automation at once, pause it for 1, 8 or 24 hours, or keep quiet hours in which nothing wakes unattended.

Pause all and quiet hours, live

What it solves

With schedules, triggers, chains and night shifts running, you need one place that stops all of it, and nights in which nothing wakes.

How it works · pick a step

The quiet hours for automations and the quiet hours for sounds are two separate settings: one keeps agents asleep, the other keeps the house quiet.

IV·9One switch1 / 5
IV·10

Forecast

What your schedules will run over the next 7 days, each with its usual time and success rate from the last 30 days.

The Tasks page, live

What it solves

Before a night or a week of unattended work, you want to see what will run, and whether the jobs you hand off come back.

How it works · pick a step

Task Analytics puts 30 days in one page: the figures, throughput, a scorecard per agent, task families and alerts, each table exportable as Markdown.

Settings · 1
SettingDefaultWhat it does
forecastDays7How many days ahead the forecast lists.
IV·10The week ahead1 / 5

Questions

Do schedules run with VS Code closed?

On Windows, yes, if you install the small headless service: it runs schedules and watch triggers with Claude while VS Code is closed. Or export a schedule as a Windows Scheduled Task.

How do I stop an empty folder from waking an expensive model?

Give the trigger a free check, such as "the folder has files", or put Gatekeeper on it. Without either, one watch once cost $70.90 in five nights of finding nothing.

What happens to a schedule in quiet hours?

It waits for the end of the window and then runs, marked deferred. A run you start yourself ignores quiet hours; the main switch still stops it.

How many runs go at once?

Two across the fleet and one per agent, by default. The rest wait, marked deferred, and start as slots free.

What if a run hits my usage limit at night?

It tries again by itself, after 40 minutes unless you set it otherwise. It can also wait for the plan's reset, or carry on on another signed-in account of the same engine.

Is Gatekeeper on by default? Is there a quick start?

Gatekeeper is off for every use; shadow shows how often it agrees with you before you turn it on, and it needs a key for Jev. For a quick start, one click sets up three examples: an inbox duty, a briefing at 09:00 and a two-agent chain.