Skip to content

Factories > Operate

The Factory control room

Open in ChatGPT ↗
Ask ChatGPT about this page
Open in Claude ↗
Ask Claude about this page
Copied!

Track work items, inspect agent runs, read factory metrics, and make the configuration changes the control room supports.

The control room is the web surface for one factory. It shows what your agents are working on, what they produced, its recorded attributed cost, and how the factory is configured. Select a factory in the sidebar to expand its pages.

Open a factory in the sidebar to reach its pages. Runs, Environments, MCPs and apps, Secrets, and Integrations sit above the factory list and cover your whole team.

PageWhat it showsWhat you do there
DashboardAutonomy, time saved, PR latency, cost per PR, pull requests, run breakdown, scorer results, and Self-improvement PRsCompare periods and find work worth investigating
ActivityWork items grouped by stageSearch, filter, open a work item, and stop one
AgentsThe factory’s agent rosterCreate and edit agents, instructions, models, runners, hosts, secrets, and MCP servers
AutomationsTriggers that start runsCreate, edit, and delete automations
RunsIndividual agent runs for this factorySearch, filter, start a run, cancel a run, and score a run
ScorersScorer definitions and resultsDefine rubrics and review classifications
BenchmarksBenchmark suites and their runsCompare harness, model, and runner configurations on fixed tasks
CodeThe factory’s definition filesBrowse and edit a Warp-managed definition
SettingsFactory identity, repos, runners, Self-improvement, integrations, and deletionChange configuration the factory owns

Activity is the stage view for the factory’s work items, rendered as one grouped list ordered by stage: Triage, Planning, Building, and Reviewing.

  • Search - Enter text in Search tasks… to match work item titles.
  • Filter - Click the filter control to add a Stage, Created by, or Created on filter. Active filters appear as chips with a running result count and a Clear action.
  • Sort - Choose Created or Updated, in either direction.

Two defaults shape what you see. Activity filters to work items you created, so a teammate’s work stays hidden until you change the Created by filter. It also shows only the four active stages. To see finished work, add a Stage filter and select Complete or Cancelled.

Filters and the open work item are stored in the URL, so a filtered view survives a reload and is shareable. Search text and sort order are not.

Click a row to open the detail pane beside the list. The pane shows the work item title, what the agent is currently doing, the stage, its origin as a Slack, Linear, or Jira ticket chip with any external references, pull request artifacts, and a collapsible details section with environment, cost, and run time. Click the clock control for Event history, a reverse-chronological timeline of the runs behind the work item. Click the short identifier in the pane header to copy the full work item ID.

Use Stop task in the pane header to cancel the work item’s current run.

A run is one agent execution. The team-level Runs page lists every run you have access to; a factory’s Runs page lists only runs executed by that factory’s agents. Both support search, sorting, and filters for source, status, execution location, creator, creation date, and agent.

Click New on a factory’s Runs page to send a prompt, with optional attachments, to that factory’s foreman agent. Click a run to open its detail pane, which shows the run timeline, the triggering prompt, cost, environment, harness, artifacts, and, for an orchestrator run, a Sub-agents tab listing its child runs. View session opens the agent’s session in a new tab.

Where a run supports it, the pane offers Stop run and Score run.

Dashboard covers a date range you choose with the presets or the date picker. Three cards head the page: Autonomy (the share of merged PRs that reached merge with no human touch beyond an approval review), Time saved (merged-PR line changes converted to approximate engineer-hours), and PR latency (median time from run kickoff through PR, review, and merge).

Below them, Cost per PR headlines the median cost of the PRs opened in the range, with By complexity and By size views and an expander for the most expensive PRs by attributed cost. It is a lower-bound estimate that can omit usage and does not match billing; see Measure and improve a factory for its limitations. Pull requests charts opened alongside merged, and Run breakdown charts total runs with a drill-down by agent type, status, source, root versus subruns, model, and harness.

The Insights panel lists Self-improvement PRs opened by your foreman, linking each one out to the code host. When scorers are available, scorer cards summarize recent classification results underneath the charts.

Agents lists the factory’s roster with search and a type filter covering Foreman, Triage, Spec, Code, Review, Monitor, and Custom. Click New to add an agent, or click a card to edit its name, description, instructions, model, runner, host, attached secrets, and MCP servers, plus the automations that target it.

Automations lists each automation with its trigger, target agent, and creation date. The editor builds triggers from a schedule (hourly, daily, weekly, or a custom cron expression) or from a GitHub, Linear, Slack, or Jira event, with per-provider filters.

Two limits apply. The agent editor sets the model, runner, and host but not a harness or environment, and the automation editor overrides no execution settings; set what these omit in the factory definition. When a factory’s definition lives in an external repository, both pages become read-only and link to the source of truth.

Code reflects how the factory’s definition is managed. For a Warp-managed definition, it opens a file tree and editor: browse the definition, edit files, and save a batch of changes as one atomic commit. Saving validates the whole tree first, and problems come back as diagnostics located to a file, line, and column in the problems drawer. An unsaved edit prompts you before you navigate away.

For a definition managed in GitHub, Code is read-only and offers Open repository instead, because pull requests there are the edit path. A live-managed factory has no source repository and says so.

Scorers defines how an LLM judge classifies completed conversations: rubric and labels, threshold, agent scope, sampling rate, judge model, and whether the scorer is active or paused. Each scorer has a Self-improvement section you enable or disable.

Benchmarks suites define an agent and its tasks, each with success criteria. The launch dialog chooses the harness, model, and runner configurations to compare, the scorers, and how many repetitions to run. Results include cost and quality with a per-task comparison; the same classification caveat as scorers applies, so read them alongside Measure and improve a factory.

Settings holds the configuration the factory owns:

  • General info - Change the factory name, avatar, and connected repos, and choose whether pull requests are authored by the agent or the run creator. The factory alias is read-only.
  • Self-improvement - Choose the Analysis model used to analyze failed runs and group recurring issues before fixes enter the factory.
  • Runners - Review, create, and edit the factory’s runners for a Warp-managed factory; a file-managed factory shows that runners/*.yaml is the source of truth and links to the repository. This section shares the Code tab’s availability.
  • Factory repo - See where the factory’s definition is hosted.
  • Factory integrations - Review the connected Slack app and cloud provider configuration.
  • Danger zone - Delete the factory. Deletion cannot be undone.

Settings respects your permissions. When you lack edit access, or the configuration is managed in an external repository, the affected controls are disabled and the page explains where to make the change.