Many agent tasks. One place to see what needs you.
Manage AI coding jobs across projects in Flame IDE. Track running and queued work, answer approvals, inspect changes, and review worktree landing decisions.
The actual Agents Manager interface with sample jobs, project groups, worktrees, and a tool approval request. Select the image to enlarge.
WHY IT HELPS
See the work waiting between started and done.
Running another agent is easy. Knowing which job is waiting for approval, which change needs review, and which task is blocked by a merge takes a different kind of workspace. Flame's Agents Manager gathers project-specific jobs and their states in one view. Move from the queue to the conversation, inspect its checkout and output, and make the next decision without losing the rest of the work.
Find the task that needs a decision
Search and filter jobs across projects, worktrees, prompts, time, and state. Distinguish work that is running from work waiting for approval or a reply, rather than treating every unfinished task as busy.
Inspect the conversation behind the status
Open a job's AI chat to read its messages and tool activity. Inspect available timing and token information alongside the project and checkout, then answer the request using the task's own context.
Connect completion to review
Keep worktree readiness and merge-blocked states visible. Inspect the resulting diff, resolve conflicts where needed, and choose an explicit landing action after the work has been validated.
HOW TO USE IT
From a job queue to a deliberate landing decision.
Give each task a clear project and outcome
Start a supported coding-agent task with the project, requirements, relevant context, and execution permissions it needs. Use an isolated worktree when the task should not edit your active checkout.
Open Agents Manager and narrow the view
Find the Agents Manager workspace, then filter by project or state and search for the task or worktree. Review queued, scheduled, running, and waiting tasks as distinct stages of work.
Handle requests with their evidence
Open a task's conversation when it needs approval or a reply. Read the requested command or tool action and the preceding context before responding. If work failed or was interrupted, inspect the reported state before deciding the next step.
Review changes before landing
Inspect the worktree and diff, run the appropriate checks, and resolve any conflict or blocked state. Use the available landing controls only for work you intend to integrate. Keep incomplete tasks separate from changes that are ready.
REQUIREMENTS & LIMITS
Know what the tool depends on.
Provider runtimes differ in tool support, permissions, and agent behavior. A shared manager does not mean every provider supports the same native worker or team controls.
A task being completed is not proof that every check passed. Inspect its reported evidence, resulting code, and validation before integrating the work.
Worktrees isolate working files. They do not automatically isolate ports, databases, secrets, or external services, and they are not security sandboxes.
Scheduled or queued work still depends on the host, configured runtime, credentials, permissions, and available resources. Keep the workstation available for the intended execution.
FAQ
AI Agent Manager: common questions
Yes. Agents Manager groups jobs by project and provides search and filters so you can follow tasks across the workspace and inspect each job's conversation and checkout.
Use the manager's task states and filters, then open the relevant conversation. Inspect the pending tool request and its context before approving or denying it.
No. Tasks retain their project and checkout context, and supported workflows can use separate Git worktrees. Choose isolation according to the task; the manager does not make every runtime resource independent.
Parallel agents describes concurrent execution. Agents Manager is the workspace for seeing job state, handling requests, inspecting output, and deciding what to review or land across those jobs.
No. Native worker observation, delegation, controls, and supported execution policies vary by runtime. Use the capabilities exposed by the configured provider rather than assuming feature parity.