Monitoring, Delivery, and Continued Iteration

Once you launch an Agent, you don't need to keep watching the full log stream. The task view tells you current progress, and when you need to provide input, approve something, or accept a deliverable.

What to monitor

In Manager, focus on:

• Working mode.
• Task status.
• Current stage.
• Sub-agent status.
• Recent events.
• Pending approvals.
• Failure reason.
• Model usage.
• Remaining budget.
• Resource access records.
• Deliverables.

Monitoring by working mode

Sub mode

Focus on:

• Whether sub-agents are all doing different things.
• Whether there's duplicate execution.
• Whether any subtask has been blocked for a long time.
• Whether results contradict each other.
• Which sub-agent is consuming the most budget.
• Whether the main Agent has aggregated the stage's results.

Swarm mode

Focus on:

• Whether tasks broken down by the lead are piling up.
• Whether any worker is duplicating work on the same task.
• Whether a worker is stuck or hasn't reported a heartbeat in a long time.
• Whether the worker count has reached the cap (up to 8, 3 by default).
• Which worker is consuming the most budget.
• Whether high-risk operations have exited the automatic flow and gone into approval.

Waterfall mode

Focus on:

• Whether the current stage is complete.
• Whether the stage output meets the conditions to enter the next stage.
• Whether any stage advanced without confirmation.
• Whether release, production, secret, and main-branch operations have been approved.
• Whether the rollback plan is complete.

Agile mode

Focus on:

• Whether this round's goal is still clear.
• Whether this round's scope has expanded.
• Whether this round's deliverables are ready to accept.
• Whether feedback flows into the next round rather than blending into the current one.
• Whether repeated changes across several rounds indicate the goal needs redefining.

Execution stages

Common stages:

Understand the goal
-> Prepare context
-> Generate a plan
-> Execute changes
-> Run verification
-> Fix issues
-> Generate delivery notes
-> Wait for acceptance
-> Optional deployment

Different tasks go through different stages. A documentation task may have no deployment stage; a deployment task may have no code implementation stage.

Viewing logs

Logs help you understand what the Agent did.

You can view:

• Current action.
• Tool call summary.
• Test results.
• Build results.
• Error reason.
• Deployment logs.
• Audit events.

Logs should never contain plaintext secrets. If you see what looks like a secret, stop the task immediately and notify the administrator.

Viewing model usage

The usage page should show:

• Consumption for this task.
• Sub-agent consumption.
• Today's consumption.
• Remaining quota.
• High-cost calls.
• Consumption from failed retries.

If consumption looks abnormal:

• Stop the task.
• Check whether it's looping on failure.
• Narrow the task's scope.
• Reduce the model cost.
• Split the task.

Viewing deliverables

Deliverables can include:

• Requirements summary.
• Technical approach.
• Code diff.
• PR.
• Test report.
• Build record.
• Deployment address.
• Release notes.
• Documentation.
• Audit summary.

Accepting the result

It's recommended to use four conclusions:

Continuing to iterate

You can keep entering things in the client:

• Modify a feature.
• Add a new feature.
• Fix failing tests.
• Adjust styling.
• Optimize performance.
• Generate documentation.
• Deploy to another environment.

When you continue iterating, Heicode reuses the existing context, resources, and audit records.

When to start a new task

It's recommended to start a new task when:

• The goal is completely different.
• The resource scope is different.
• A different approver is needed.
• A different budget is needed.
• The previous task has already been delivered and archived.

When to stop a task

It's recommended to stop a task when:

• The direction is wrong.
• Permissions are misconfigured.
• The budget is out of control.
• There's a resource authorization error.
• A security risk has appeared.
• The task can't continue.

After stopping, check the audit trail and temporary resources first, then decide whether to duplicate the task and start over.