Activity

Activity is the list of the work your agent did for you. Not a chat log and not a raw event stream: one row per task, titled in plain words, with the outcome underneath. Open a task to see what it did step by step; open a step to see exactly what the tool was given and what it returned.

It answers "what has my agent been doing?" at a glance — the research it ran while you were away, the routine that fired at 8:00, the email it sent after you approved it, the goal it is still working through.

[!NOTE] Activity lives entirely on your machine, in your agent's own home. Every agent has its own Activity. Nothing is uploaded, and only someone who can open a conversation can see the work done in it.

Where to find it

All three read the same journal from the agent itself, so a task looks the same on every device and the unread dot clears everywhere once you have looked.

Work, not conversation

Activity only records work. Saying hello, asking a quick question, or thanking your agent is conversation: it stays in the chat and never appears in Activity — not even its words are written down.

A turn becomes a task when:

• a routine or a goal started it — those are always tasks; or
• your agent took a step that is work — searched or read the web, opened or wrote a file, ran a command, used a connected app, generated an image or a video, sent a message, asked another agent.

Some tools are part of talking, not work, and don't make a task on their own: recalling from its memory or your past conversations, asking you a question, drafting a plan for your approval, reading its own skills. A turn that only did those is still conversation.

A long piece of writing with no tools at all — a cover letter, a plan, a report of 1,500 characters or more — is a candidate: your agent's model decides whether it was real work. Messages that reach your agent through Telegram, Slack and other channels follow exactly the same rules as the app.

Your agent's model also gets the last word in the other direction. When it writes the summary it may judge that a turn which used a tool was really just conversation — a single quick lookup to answer a passing question — and that turn is left out. It never does this for a routine or a goal.

One task, several turns

Real work often takes more than one message. Activity keeps it together:

• Goals. Every turn of a standing goal belongs to one task from the start, however many turns it takes.
• Follow-ups. If you carry on recent work in the same conversation — "now fix the second issue", "look again" — within 30 minutes of the earlier work ending, your agent's model judges whether the new turn continues it. If it does, the steps join the earlier task and its title and summary are rewritten to cover the whole thing. A new subject starts a new task.
• Routines. Each run of a routine is its own task, named after the routine.

A task keeps its place in the list (where it started). While a follow-up is still running it can briefly appear as its own row; it folds into the earlier task a few seconds after it ends, once the summary is written.

What a row shows

 ◌  “find me cheap flights to Rome for the 12th”        now
    • Searching for “Rome flights November”

 ◎  Compare flights to Rome                             14:02
    Found three fares under 200 euros

• The icon. Three calm drawings cover everything: a globe in a magnifier for research on the web, a clock with a turning arrow for a routine, and four nodes on a chain — a task's steps — for everything else. While a task runs, the chain's nodes fill one after another.
• The title. Written by your agent's own model once the task ends, in the language you spoke — never copied from your message. Until it arrives, your request stands in, quoted and in italics, so you always know which task is which.
• The line under it. The outcome once there is one. While the task runs, what it is doing right now — Searching for …, Reading …, Running … — with a small dot.
• Status. Done tasks stay quiet. Anything else is labelled: running, waiting for you, stopped, failed, blocked, interrupted.
• The unread dot. Tasks that need a look — failed, blocked, waiting on you, or cut short — carry a dot until you have opened Activity.

The list refreshes itself: every couple of seconds while something is running, otherwise every ten seconds.

Statuses

Opening a task

A task opens with:

• its title, outcome and current status — live while it runs, with new steps appearing as they happen;
• a short summary in your agent's own voice;
• the steps it took, each with a one- or two-sentence note on what it found or did;
• the approvals and questions it raised, and how each was answered;
• when it started, how long it actually worked (time spent waiting on you is not counted), who started it (you, a routine, a goal, a channel), the model and the tokens it used;
• Open conversation, to jump to the chat where it happened.

Opening a step

Open any step to see it in full — exactly as the tool panel in the chat shows it:

• what the tool was given (the search query, the file, the command, the request), and
• what it returned, laid out the same way the chat lays it out: search hits as results, files as files, command output as output.

These details are not copied into Activity. They are read back from the conversation itself when you open the step, so they are only as available as the conversation is. A step from a conversation you have deleted, or recorded by an older version, simply says the details are no longer kept.

Privacy

• Local. The journal is plain files in your agent's home, one per month: ~/.flowly/activity/<YYYY-MM>.jsonl for your main agent and ~/.flowly/profiles/<name>/activity/ for each other agent. Nothing is sent anywhere.
• No conversation. A turn that is only conversation writes nothing at all.
• No arguments or results. A step records its tool, a short target (the search query, a file's name, a website's host, a program's name — never its arguments, and never anything that looks like a key or token), whether it worked and how long it took. What a tool was given and returned stays in the conversation.
• Same access as the chat. A task, and each of its steps, can only be read by someone who may open the conversation it happened in.

Summaries and their cost

Titles, outcomes, summaries and step notes are written by your agent's own model, in one short request per task after it ends — never while it is replying. The request carries a compact account of the task (what was asked, each step and the start of its result, the start of the reply), never the whole conversation. The judgements — was this work?, does it continue earlier work? — ride in the same request at no extra cost. Conversation is never sent, because it is never recorded.

You can switch summaries off. Tasks are still recorded; each turn that did work is then its own task, its title is the routine's name or your request, and borderline turns (long writing with no tools) are left out.

Settings

In ~/.flowly/config.json (or the agent's own profile):

{
  "activity": {
    "summaries": true,
    "retentionDays": 90
  }
}

Activity and the audit log

The audit log is the forensic record: every model call and tool call, for debugging and review. Activity is the owner's view: only the work, put together into tasks, in plain words. They are independent — turning one off never affects the other.

Related

• Standing goals
• Cron
• Sandbox and approvals
• Audit log