Put your routineson a schedule

Finish a job in chat, then say "turn what we just did into a workflow". Plexon reads the conversation, the tools it called and what came back, and drafts a routine that keeps the steps that decided the outcome and drops what was specific to that one case. A date, a file name or a customer from the chat becomes an input or a setting, never a fixed value. You can also name past chats, or past runs of another routine, up to five together. The draft arrives switched off, like every draft.

• Learning mode: ask for a routine that learns, and each agent step writes a few short notes while it works. Which step settled the job, what it opened that could not have changed the answer, which source worked when another did not
• The notes are about the kind of job, not the one case. They are never shown as facts and never become memories
• A draft from past work goes through the same checks as one you described, and lands switched off
• Ultra for a whole routine, or for the one step that needs it. The switch is on the routine, the override on the step under Advanced

A routine reads top to bottom as plain cards, one line each, with the values it reads named in words: a condition as a sentence, a wait as "3 days", a schedule as "Runs every weekday at 09:00". Select a step and its panel opens on the fields you fill in. Everything else (the connector, the ids, the time limit and retries, the raw JSON) waits in one Advanced fold at the bottom.

• A message or prompt that mixes text and values shows each value as a pill in the flow of the text. Add one with + value or by typing {{, remove it with Backspace, and open a long prompt in a larger editor with the values listed beside it
• A rule on a condition is built from pills and menus: pick a value, choose a comparison that fits it (text, number, yes or no, list, date), type or pick the other side. Rows group under All of these or Any of these, groups nest, and the same rule shows as code with one click
• A schedule that runs more than once a day lists its other times under Also at, next to the first one
• An agent that asks a few questions shows them as fields, and the prompt box becomes whatever else you want to tell it
• The What is on the screen? card takes its picture from the browser step above it, so nobody types the path
• Every timeout is a number and a unit, minutes to weeks, and the Run dialog asks for each input by a plain name, with a dropdown where the answer is a code
• Nothing sits behind a mode. The Advanced fold opens on its own when a field in it has an error or already holds a value
• Every routine reads as one paragraph, written from its steps: on its card, in its panel, and above each run’s report
• Notes live in a list beside the canvas, and a note can point at one step

Choose when the first step should run. A trigger is the time or event that starts it.

• A schedule: every morning, each Monday, several times a day, or at the interval you choose
• A new item in an app, such as a Jira ticket, Asana task, or HubSpot contact
• A regular check of an app for new items, such as new mail in Gmail or Outlook
• A new article in a news feed
• A Telegram or WhatsApp message that matches your filter
• New activity on a GitHub project you already watch in Plexon
• A signal sent by another app
• Something happening in Plexon, such as a finished reply
• The Run now button, with any details you need to enter first

A routine can act in your apps, ask AI to read or write something, and decide which step comes next. Use AI only for the steps that need it. Sending a message or moving a file can run directly.

• Pull an invoice number from an email, draft a reply, or summarize a report
• Give a step to an agent, a specialist assistant, or a persona, a saved role with its own instructions and tools
• Repeat an action for each item in a list, such as every row in a spreadsheet
• Choose between steps based on a condition, such as whether an email includes a phone number
• Have a separate judge read the evidence before a routine accepts that a piece of work is finished
• Put a closed question to a fast decision instead of the AI: is this an order, how urgent is it, which of these categories fits. Several questions about one result ride one call, and the routine branches on whichever answer came back
• Wait for a reply, pause for your approval, or send a notification to your desktop, Telegram, Slack or WhatsApp
• Show the desktop notification only when none of the chat apps you picked took the message. Add a link and text to copy, and clicking it opens a card with Copy text and Open the page
• Add a random extra to a wait, so a daily run does not start at the same minute every day
• Give an agent step the websites it may open, and it works them in a hidden browser window on your signed-in accounts, with Show and Stop in the Workflows panel
• Answer a delivered report: reply to the notification on Telegram or Slack, or click it on the desktop, and a chat opens that starts from that report
• Open one of your own chats and send it the finished work, in Plan mode when you want a plan rather than edits
• Build a table, calculate totals, or tidy names and phone numbers without asking AI

A Decide step hands a result to Jev, the fast decision model, with up to eight questions about it, and gets every answer back in well under a second with a probability on each. It can branch on one answer, with each option its own edge. It can keep every answer for later steps to read. It can go through a list and keep the items that pass a yes-or-no question, or pick one item out of a list. When an answer is not sure enough, the routine can take a Not sure edge, to an approval step say, rather than guess. A Decide step can also look at a picture: point it at the screenshot an earlier browser step stopped on and ask a yes-or-no about it, or ask it to find something and get back where it sits, as the four sides of a box, in fractions from the top left. If the earlier step left no screenshot, the step takes its fallback. It costs about 1,000 to 1,800 input tokens on a run that has a picture, and a switch in Settings turns it off.

• Every answer and its probability stay on the run, and a later step reads them by name
• Replay: after editing a routine, ask Plexon to replay its recent runs against the edit. Nothing is called and nothing is sent. It re-reads the answers each run kept under the new thresholds and edges, runs the code steps over the old results, and reports which runs would have taken a different path
• A question the old runs never asked is reported as unanswerable, not guessed. Ask for a live replay and only those questions are asked now, with a cap on how many decisions one replay may spend
• Each routine keeps its last 200 runs to replay against, and can be told to keep up to 1,000
• With Jev off or unreachable, the step takes the fallback edge its author named and the run goes on

An agent step that fills in named fields, say a total and the link it came from, can be asked to back each one up. The agent quotes the exact excerpt from the page, file or tool result that establishes each value, and Jev reads each value next to its excerpt and says whether the excerpt supports it. A field with no support goes back to the same agent, which quotes better support or reports it as unknown. After the rounds you allow, up to three, whatever is still unsupported is dropped from the result, or the step fails if you prefer that. What stays is backed by something the agent read.

• Choose which fields to check, or check them all, up to eight
• Each round is one fast decision, not a chat turn, and the agent is sent back with one message naming only the fields that failed
• The run shows which fields were kept, which were dropped, and the support each one had
• With Jev off or unreachable the step keeps its result exactly as extracted and says the check did not run

Each step has a form with named fields and examples. Choose information from an earlier step, such as the address found in an email. Under every field a line says what it comes to with the last run’s data: the message as it would read, a condition as Yes or No, a loop as how many times it would repeat. Test step runs one step for real or as a dry run; a test send previews the message without sending it.

• A value the data does not have is named in plain words, never as a path
• The preview only reads. It never runs a tool, asks the AI, or sends anything
• Browse for a file or folder using the usual picker on your computer
• See matching files as you fill in a search
• Use the information menu to pick a result from an earlier step
• Errors appear beside the field that needs attention, with a suggested fix when available
• Review the next scheduled runs before you save

An outreach routine can remember who it contacted, which channel it used, and what was said. Replies can continue the same conversation across email and WhatsApp. Open that conversation in the sidebar when you want to take over.

• Find contacts by status and read their message history
• Continue a previous exchange instead of starting a new conversation
• Automation pauses when you take over and waits until you resume it
• Contact records are saved on your computer; sending messages uses the connected service

Open the Runs tab to see each routine and its recent runs. Each run opens with its report: the numbers it found as tiles and its message, then every step in plain words, how long it took, and what it cost.

• Line up a routine's runs in a progress table with a total for each number
• Save one run, or a whole week of them, as a PDF report
• Click a routine's notification to open that run's report
• The first failure tells you, with a fix: one notice, one button (Sign in again, Install, Open the step, Open settings). The same cause stays quiet until you fix it
• See when each schedule runs next, and stop a run that is still going
• Plexon keeps the computer awake while a schedule is set, as long as it is plugged in. A closed lid still puts it to sleep
• A run whose time passed while the computer slept is skipped and the card says so, or runs once at wake with Catch up on
• See the last 24 hours, the last 7 days, or the full history
• Ask in chat how a routine's runs went and the answer counts the runs that handed a case back instead of deciding it, and the fast decisions each one spent
• Runs waiting for your approval stay easy to find
• Clear finished runs individually, by routine, or once they are 7 or 30 days old
• A running job must be cancelled before you can delete it

Plexon includes a routine for jobs too large for one answer. Give it an objective and a limit on how many rounds it can run. Each round reads the previous work, tackles the next part, and reports what changed.

• Set the round limit and spending budget before it starts
• Read each round, stop the job, or follow its progress in Workflows
• The routine stops when the work is complete, it needs help, or it reaches your limit

Include a workflow in a persona or share it through a marketplace, a collection of tools and instructions people can install. Your teammates can use the same starting point and adjust it to their work.

• Include routines with the saved role that uses them
• Share through a marketplace you or your team manage
• Choose which projects can use an installed routine