Guides

Run an agent on its computer

This guide runs Tablua's own agent, the one that builds small apps, on a Moss computer, with all three models.

Important

Today a full agent run needs two small Lua modules from uspx, Tablua's post between agents, which is not public yet. Until it is, the harness itself runs from a public clone (see the Quickstart and Embed the harness in Lua), and this guide describes the run for when uspx is published or for those who have it. Status and roadmap tracks it.

#What you need

• Elixir 1.15 or newer, with Erlang/OTP
• API keys for the three models:

#1. Set up

git clone https://github.com/OpenRelationship/tablua.git
cd tablua
mix setup
mix moss.look          # fetches the browser's page-reading module

#2. Start a computer

iex -S mix

Moss.Computer.run("plants", "ls")

A computer named plants now exists, with its own SQLite file. Its name can be lower-case letters, digits and dashes.

#3. Give the agent a task

ask = "I'd like a little app for my house plants. It should list each plant with when I last " <>
      "watered it, let me add a plant, and let me mark one as watered today."

Moss.Computer.Agent.run("plants", ask)

The agent takes steps until it finishes, stops, or needs you. It returns a map:

%{outcome: "waiting", why: "waiting for ...", steps: 2, counts: %{"jev" => 2, "mercury" => 1, ...}}

outcome is one of done, blocked, waiting, stopped or error.

#4. Play the person's part

The agent never agrees to its own feature or approves its own publish. Those are yours.

Agree to the feature it wrote, after reading it:

Moss.Computer.run("plants", "cat features/plants.feature").out |> IO.puts()
Moss.Computer.agree("plants", "features/plants.feature")

Then run the agent again. It picks up where it stopped:

Moss.Computer.Agent.run("plants", ask)

Say yes to publishing when it asks. The log shows the request the agent made; answer it with the same line:

Moss.Computer.answer("plants", line, true)

To automate the person's part, for example in a test, pass a function as between:. It is called with the computer's id after every step.

#5. Read what it did

Every step is in the computer's file as Tablua rows. Locally, a computer's file is priv/work/computers/<id>.sqlite:

sqlite3 priv/work/computers/plants.sqlite "select n, chosen, by from tablua_decision"

The queries in Read an agent's file with SQL work on it directly.

#Options

Moss.Computer.Agent.run/3 takes these options:

#Next

• The step loop
• Turn learning on