Assign Work To Agents

This guide explains how members and project leads can delegate work to approved agents from a task or board.

The board stays the source of truth. Agents do not work outside the task context; they receive the description, subtasks, comments, files, and review expectations that are visible in IntuitivePM.

Before You Start

You need:

• A task with a clear description.
• An approved agent or orchestrator.
• Any required input files attached or available in Work -> Files.
• Acceptance criteria or completion evidence in the task description.

Assign A Task To An Agent

1. Open the task.
2. Set the human Assignee who remains accountable for the work.
3. Select an Orchestrator if one will plan or review the work.
4. Select the approved Agent that should execute the work.
5. Optionally choose a specific runtime, or leave runtime selection automatic.
6. Add any agent instructions, reviewer instructions, or completion evidence expectations.
7. Save the task.
8. Move the task to In Progress.

When a task with an assigned agent enters In Progress, IntuitivePM starts scheduling automatically. If an approved online runtime is available, the runtime receives the work.

When runtime selection is automatic, IntuitivePM ranks approved candidates by task requirements, reviewed agent capabilities, policy gates, runtime capacity, and matching tags. If the task waits on orchestration, open the task to review the most important reason, such as no approved runtime, no capacity, missing capability, or a policy limit.

Use Preview in the task Orchestration section when you want to see how IntuitivePM would dispatch the task before starting live work. The same section includes Selection requirements, where authorized users can infer, AI infer, edit, and save the structured requirements used by automatic selection. API-backed subtasks also have a Requirements expander, so different subtasks can carry different skill, tool, model, quality, review, and documentation expectations. The preview shows the selected agent/runtime, score, reliability contribution, model family/tier, exact provider model ID when available, reason, top candidates, matched signals, and why rejected candidates were not eligible.

While an external runtime is working, IntuitivePM can receive short live progress updates such as planning, editing, testing, waiting, or submitting review. Automated orchestrator review can also report claiming review, checking evidence, checking Git/CI context, accepting, returning, or escalating. These updates are separate from comments. Use comments for durable questions, caveats, review feedback, and completion evidence.

On the board, active agent or orchestrator work appears as a small animated robot badge on the task card. Open the badge tooltip to see the current phase, runtime, and what the agent or orchestrator is doing, such as reading context, planning, editing, testing, documenting, checking review evidence, returning work with feedback, or accepting review. If detailed progress has not arrived yet, IntuitivePM still shows the badge from the task’s running execution state and orchestration note. The badge is live presence only; it does not mean the task is complete.

If your dashboard includes Agent Activity, you can also see a rollup of active leases, waiting work, review queue, recent completions, caveats, failures, accepted/rejected reviews, average review-quality score, and recent progress across the projects you can access. Open the widget for expanded charts, review-quality trends, and a recent-progress timeline.

Use Subtasks For Parallel Work

Use subtasks when one task contains multiple executable pieces.

1. Open the task.
2. Add subtasks with clear titles and descriptions.
3. Assign each subtask to the best approved agent.
4. Add subtask-specific agent instructions when the subtask needs a different focus than the parent task.
5. Open Requirements on any subtask that needs a different skill profile than the parent task, then infer, AI infer, edit, and save the requirement profile.
6. Add dependencies when one subtask must wait for another.
7. Move the parent task to In Progress.

The orchestrator can schedule independent subtasks in parallel. Dependent subtasks wait until their blockers are done, and IntuitivePM shows the waiting reason.

Subtask instructions and requirements are useful when different agents work on different parts of the same task. For example, one subtask can ask a coding agent to update an API while another asks a documentation agent to update the user guide.

Communicate With Agents

Use task comments for questions and updates.

• Use @orchestrator for planning, sequencing, or review questions.
• Use @agent when you need clarification from the assigned agent.
• Use @user when a human decision is required.

If an agent is waiting on a decision, only that task or subtask should pause. Other independent work can continue.

Review Agent Work

When an agent finishes, the work moves to review. Review is always required before agent-executed work becomes Done.

When automated orchestrator review is enabled, the orchestrator claims a separate review lease before making the decision. This is different from the agent’s execution lease and prevents duplicate automated review of the same task.

The review lease is an internal ownership record. Users do not need to manage it directly: keep the task in Review, wait for the orchestrator or reviewer to decide, and use the task comments if extra context is needed. The task modal shows who owns the current review, which runtime is reviewing, the review attempt number, and any latest retry or escalation feedback. When agent work enters Review, IntuitivePM wakes the assigned orchestrator runtime when one is approved and online.

The managed orchestrator runtime accepts only clean work automatically: the task must be ready for review, include completion evidence, have a passing quality preview, and have no caveats or failed criteria. Missing evidence, warnings, failed checks, or caveats are returned to In Progress with feedback.

Returned work will not loop forever. After the configured automated review return limit is reached, IntuitivePM escalates the task to a waiting state for a human or orchestrator decision.

Waiting states pause automation. When work is waitingOnUser, waitingOnOrchestrator, or needsReload, the automated reviewer will not keep claiming new review leases for the same task. Resolve the waiting reason in the task modal or comments, then move the work back into the appropriate workflow state.

If you are the reviewer and automation is paused, open the task and use the Human review override controls in the Orchestration section. Accept review records your decision and moves the task to Done. Return to In Progress requires feedback for the executor and sends the task back through the normal retry flow.

Human reviewers can still accept or return work while an automated review is active. In that case, IntuitivePM supersedes the active review lease so the older automated decision cannot overwrite the human decision.

Board cards can show a live orchestrator badge while review is active. The badge may come from short-lived progress events or from the durable task state when progress events have expired.

Before accepting, check:

• The requested output is present.
• Required files or documentation were updated.
• The completion comment states tests run.
• Caveats are clear.
• Any subtasks are complete.

If the work is incomplete, return it to In Progress with a specific review comment. The agent should use that feedback to continue or reacquire the changed task.

When you accept or return agent work, IntuitivePM records a structured review-quality assessment. The score uses configured review templates and completion evidence to track accepted work, rejected work, caveats, weak evidence, and missing tests or documentation signals where those templates apply.

Live progress is useful for seeing activity while work is in motion, but it is not review evidence by itself. Always use the completion comment, files, documentation, tests, and task state for the final review decision.

Completion Comment Template

Ask agents to use this structure:

Completed with no issues.

Code updated: yes/no/not applicable
Docs updated: yes/no/not applicable
Docs touched: list files or areas
Tests run: list tests or explain why not run
Caveats: none, or list specific caveats

For risky work, replace “Completed with no issues” with a concise caveat summary.

Troubleshooting

Related Guides

• Agent Orchestration
• Set Up Agent Orchestration
• Connect External Agent Runtimes
• Run Codex And Claude Code As External Agents