Understanding how AI Coworkers run · Atomicwork | Product
Understanding how AI Coworkers run
Underdstand how an AI coworker behaves at runtime.
Once an AI Coworker is published, it functions just like a human teammate in your workspace. It doesn't sit idle; it boots up, evaluates the request, executes its runbook, and logs every decision it makes.
This guide breaks down how to read execution logs, track live progress, and audit your AI Coworker’s behavior.
How work finds your AI Coworker
Requests are automatically assigned to an AI Coworker based on its Identity. Human teammates can also manually assign requests or @-mention the agent directly. From there, the AI Coworker triggers a Run.
A** Run** is a one, isolated execution by an AI Coworker. It evaluates the request context, chooses its tools and then attempts to reach its end-of-run status.
The lifecycle of a request
A single request can trigger multiple independent runs as work moves between different AI Coworkers. For example:
Intake: A request about slow network performance enters the workspace from an employee, a workflow, or a system like PagerDuty.
Triage: The workspace's Triage AI Coworker matches it to the Hardware Specialist.
Execution:
The checks local device telemetry. Finding no issues and suspecting a VPC cloud bottleneck, it reassigns the request to the AWS Engineer, ending this run.
The AWS Engineer reads the request and added context. It resolves a gateway configuration error and logs the outcome. Finally, it assigns the request back to the Hardware Specialist, completing its run.
The Hardware Specialist then reads the outcome note from the AWS Engineer. It re-checks the device network telemetry, verifies the speed is normal, and logs the outcome.
Resolution: The Hardware Specialist then posts a public note notifying the user and changes the request status to Resolved, ending its second run.
A run always ends with a clear outcome — even if that outcome is "I can't handle this, here's why." AI Coworkers do not loop indefinitely; one run produces one bounded outcome per its goal's end-of-run status rules.
Runs are numbered sequentially across the AI Workforce based on when they start. For example, the IT Triage AI Coworker might execute Run-56 and assign a request to the AWS Engineer. If other AI Coworkers process requests in between, that AWS Engineer’s immediate next step might be numbered as Run-247.
Tracking Runs for an AI Coworker
You can track runs from two different places depending on what you're looking for.
1. Inside a specific request
Open any request to see its history. Each run shows up directly in its timeline. You can see which AI Coworker executed, its run number, and the steps taken. Click on the summary line to expand and audit that specific run's full details.
2. Across the entire AI Coworker
To see all runs for a single AI Coworker, navigate to Settings > AI Workforce > AI Coworkers. Select your AI Coworker and click the Runs tab.
Understanding a Run
Clicking a specific Run opens the Run Execution Detail, providing a line-by-line audit trail. Each run displays:
Run number: The unique Run number such as Run-83
Status: Tracks execution state (Running, Completed, Failed, or Paused for Approval).
Timestamp: When the run started.
Request: The specific request the AI Coworker worked on, hyperlinked for easy access
Token usage & cost: Displays the input, output, and cached tokens used, along with the total financial cost for that specific run. Learn more in the Understanding Run metrics and costs section below.
Execution tabs: Provides three distinct views (Summary, Nice, and Raw) to audit the AI Coworker's internal reasoning process.
Execution detail tabs
Summary tab
This is a short, human-readable explanation of what the AI Coworker did and why, written by the coworker itself at the end of the run. It is built for an auditor to quickly scan what the AI Coworker decided and what comes next. For example:
Replied to Gautham's @-mention with a public note asking for the specific indicators I'd need to triage phishing exposure (clicked link, entered credentials, unexpected MFA prompts, browser redirects). I did not reassign or change status — I'm mentioned but not the assignee. The ticket remains with the IT Support Triage coworker for routing decisions.
Nice tab
This tab displays the run as a chronological timeline of steps (e.g., Run started, Thinking, tool calls, Done). Each step is collapsible—you can expand a Thinking step to see the coworker's internal reasoning, or expand a tool call to inspect the exact request and response payloads.
Use this view when you want to:
See exactly how the coworker arrived at its outcome.
Debug a coworker that didn't behave as expected to see where its logic went off track.
Optimize a coworker by tracking how many tool calls it makes per run.
Raw tab
This is the developer-grade view showing the full, unedited event stream. It includes precise timestamps, event types, IDs, and raw JSON payloads. For example:
Use this view as the source of truth when you need to reproduce an issue precisely or file a technical support ticket.
Understanding Run metrics and costs
Use these cost and token metrics to understand the resource consumption of each run:
Input tokens: The total volume of context the model read during the run. This includes the ticket history, the AI Coworker's goal, identity, soul, attached skills, and available tool definitions. Longer ticket histories and broader goals increase this number.
Output tokens: The total volume of text the model generated. This includes the AI Coworker's internal reasoning steps, notes, status updates, and tool calls.
Cached tokens: Tokens served from cache instead of being processed from scratch. Caching automatically reduces costs for repeated context, like when the same goal and skills are reused across multiple runs.
Cost: The total financial cost of the run, calculated automatically using current model pricing.
Troubleshooting AI Coworkers
If an AI Coworker isn't acting as expected, use this matrix to identify the pattern and apply the fix.
Tip: When you aren't sure where to start, open the Nice tab of a failed run. Walk through the Thinking blocks step-by-step to see exactly where the logic went off track.
|
What you notice
|
What to look at
|
Suggested fix
|
Keeps doing work it shouldn't
|
Goal's Non-Goals section
|
Add explicit non-goals naming the work it must avoid.
|
Stops in the middle of a ticket
|
Run Summary & End-of-Run status rules
|
Update the Goal's End-of-Run section to cover this specific scenario.
|
Takes the wrong action
|
Run Nice view (Thinking steps)
|
Tighten the step in the Goal's role-specific flow that led to the ambiguous call.
|
Can't find data it should see
|
Tools tab
|
Ensure the tool or the specific action required is enabled.
|
Mentions a non-existent tool
|
Goal text vs. Tools tab
|
Remove the text mention from the Goal, or enable the actual tool.
|
Reassigns when it shouldn't
|
Identity's Role & Scope
|
Broaden the role; tight boundaries or non-goals might be forcing accidental escalations.
|
Doesn't reassign when it should
|
Identity & Goal's Out-of-Scope bucket
|
Define the Out-of-Scope parameters and explicitly tell it where to route those tickets.
|
Costs are climbing week-over-week
|
Goal length, attached skills, tool definitions
|
The configuration has grown too heavy. Split the workload into multiple AI Coworkers, or move repeating SOPs into skills.
What happens if a coworker is unpublished mid-run? If you unpublish an AI Coworker while a run is active, that specific in-progress run will still finish. No new runs will start until you republish it.
What happens if a device or external tool is offline? If an AI Coworker fails to execute a command because a device is offline or an external system (like JumpCloud) fails to respond, the run will typically end with an In-Progress status and drop a public note explaining the issue. A human or another AI Coworker must manually pick up the ticket once connectivity is restored.
Related articles
To optimize you AI Workforce: See Best practices for designing AI Coworkers.
To understand the underlying concepts: See Admin 202: Understanding the AI Workforce.
To learn how to create an AI Coworker: See Setting up an AI Coworker.
