Overview and performance

The Overview page

Overview, beside Tasks and Schedules in the sidebar, shows every active project in one table, in the sidebar’s order. A project filter at the top narrows it to one project.

• Columns: Agent; Waiting on you (unanswered questions, unread replies); Tasks (open, working, blocked, stale: open and unchanged for over 7 days); Tokens (the last 5 hours and the last 7 days); Now (processor and memory).
• Rows: a first row with the totals, then each project (its name, status, its remote computer when it runs on one, how many agents are working, idle and stopped), then one row per agent.
• Signals in place: what needs attention shows in its own row. A blocked task or question waiting over an hour shows its wait in amber (“on you 6 h”); an open task unchanged for over 7 days is amber (“unchanged 12 d”); a task marked working while its agent has done nothing for over 10 minutes shows “idle 25 min”. Red is kept for a crash loop.
• Easy to read: a zero shows as a faint dash; only what needs attention is coloured.
• Tapping an agent’s row opens its detail page, a project’s row the project’s page; a count opens what it counts.
• Clicking a column’s heading sorts the agents within each project by it.
• The headings and the Agent column stay in place while the table scrolls; on a phone it scrolls sideways.
• It updates live.

Where tokens go

Tokens by task ranks the last 7 days’ tasks by the tokens they used, each with its type, a bar and its figure, the rest in one “Other N tasks” line. A task that used at least twice the usual tokens for its type is marked (“3× usual”). Tokens per task type lists each type’s tasks finished in the last 30 days, with their median and total tokens.

Agent and project detail pages

• Agent detail page: its project above its name, its state and sub-agents, its setup, and Open chat; then its open tasks and questions, the sub-agents running, Tokens by task, its problems, and a Processes card over 24 hours, 7 days or 30 days with its busiest programs. Its times: last worked, last reply, your last message, active schedules.
• Project detail page: the project’s name with Open chats, its state and where it runs, its times (including the last deploy), its machine warning, its Agents in one table, Running now with its processor and memory and a Processes card over all its agents, its open Work, and its Tokens.

Problems

Each agent measures its processes, and only programs with something wrong show, each with what to do:

• Left running: processes idle for over 30 minutes still holding memory, such as browsers left by finished tests.
• Crash loop: started 10 times or more in an hour, each living under a minute.
• Runs often: a test or build run over 20 times in a day.
• Heavy: over one core for 10 minutes or more, or over a quarter of the memory.

Nothing is stopped from the page: you ask the agent to reduce its load.

Meters

Each project row in the sidebar carries a meter of that project’s own agents and everything they started (tools, commands, sub-agents, background jobs). Other processes don’t move it.

• Its processor use is a share of the cores the project may use, its memory a share of its memory limit (or the machine’s).
• Low, it shows only while you point at the row; it turns amber at 80% of its cores or 85% of its memory, and red at 95% processor for 30 seconds or 95% memory.
• Its tooltip names the figures and the agent using the most; clicking it opens the project’s detail page.

Machine warnings

At the top of the Overview (for every machine) and of a project’s detail page (for its own machine), one line for each strained machine, and nothing when every machine is fine:

• this server: amber when tasks had to wait for the processor or the disk much of the last 10 seconds, or under 10% of memory is free; red when everything stalled or under 5% is free;
• a project’s container: amber from 80% of its cores or 85% of its memory;
• a remote computer: under 10% of its memory free, or busy at all its cores.

The line says what strains it (“This server is short of memory”, “This server is busy”), short figures (“15 of 16 GB used”), and what uses the most. Ask agents to reduce load at its end asks the agents of the projects behind it to stop what nothing needs any more and run heavy work a lighter way, then report what they changed. On this server it is offered only when the projects use at least a quarter of what is in use. The same button is on the Processes cards.

Token figures

An agent’s token figures are its turns’ use, sub-agents and compactions included. A turn’s use goes to the tasks being worked on in it, to the action or workflow run it took in (“Action runs”), or else to “No task”. They are kept for 90 days, and trashing, purging or renaming never loses or changes them.