Show It Once: Turn a Repetitive Task into a Weekend Agent

The task should already have a script in your head

The best weekend agent starts with a back-office chore you can explain to a new hire in a few sentences. If you need a whiteboard, a policy manual, or a meeting to describe it, choose a smaller task. Most side projects stay side projects, and that is okay.

Caddi creates production agents from a one-time narrated screen-share demonstration of a back-office task and lets users revise the automation using ordinary English. The agent should do one job, in one place, with one obvious end state.

Say you move files, draft a status note, or clean up a spreadsheet before a report. Pick a task that repeats often enough to matter, but is narrow enough that a bad run costs an hour, not a week. Name the task plainly: a cleanup, a summary, a triage pass.

Caddi has a Discover feature that scans eight connected tools for repetitive manual work and operates with read-only access. A read-only scan is a useful way to find the task before you promise it to an agent.

The checklist below keeps the build small: show the work once, keep the instructions in plain English, and check the evidence before you trust it.

A demo can be the whole specification

• Choose one repetitive task you can finish in a single sitting. It should have a clear start, a clear end, and a result you can check quickly.
• Narrate a screen-share demo while you do the task once. Say what you are looking for, why you click, what counts as done, and what you would do if the data looks odd. You are done when a stranger could repeat the job from the recording without asking you a question.
• Write the agent's job in plain English, not in a prompt that sounds like code. Name the trigger, the inputs, the allowed actions, and the output. The result is a short paragraph you could paste into a team channel and still understand.
• Scope permissions before the first run. Give the agent only the tools, folders, and actions it needs, and keep destructive actions out of reach. The usual failure is giving the agent broad access before it has earned a clean run. Done means a permission list you can read aloud without flinching.
• Review run logs before leaving it running. Check what it touched, what it skipped, and what it could not decide. The log should be short enough to scan in a coffee break and trust enough to let it run overnight.

Plain English is the maintenance interface

The maintenance interface should be the way you already talk. If the agent needs a new rule, say it the way you would tell a coworker. Caddi's Loop Studio can learn a process either from a screen-share demonstration or from a description of the desired outcome, then generate a plan and build the agent. Read the plan before you let the agent run.

The trust part is the log. A run should show what it did, what it skipped, and where it stopped. Caddi differentiates itself from RPA, or robotic process automation, and workflow builders by pairing AI-based reasoning with deterministic execution, while recording each run and limiting permissions. For a weekend project, that matters more than a fancy dashboard.

Start with read-only access where you can. If the agent only needs to find files, summarize notes, or draft a reply, do not give it write access until it has earned it. A clean run is a clear record: the agent did what the log says, and the log says what you expected.

A clean log is what earns the next run

Run the agent for a few cycles before you stop checking it. If it handles the normal case, test one edge case: a missing file, a duplicate row, or an unexpected value. You are not trying to prove the agent is smart. You are trying to prove it is boring.

A boring agent does the same thing, in the same order, with the same evidence. When the log looks like a receipt, not a mystery, you can let it run while you get back to your day job.