All articles
AI Agent Limitations and Troubleshooting
TL;DR
AI Agents have key limitations: they can't delete items, edit descriptions, access attachments, external data, or workload metrics, and are unreliable at math. Behavior is non-deterministic and limited to assigned scope. To troubleshoot: ensure the agent is toggled ON and @mentioned in a location, verify triggers match actual events, check the Activity Dashboard for AI reasoning, test in the Playground, and list team members by name for accurate assignments. Note limits on email, Slack, Teams, subitems (max 1,000), and date/time fields.
| Availability: Business, Pinnacle, Apex. ; Unavailability: Free, Team; |
- Overview
- Behavioural Limitations
- Actions AI Agents Cannot Perform
- Data AI Agents Cannot Access
- Consolidated Limits Reference
- Troubleshooting Decision Tree
- What’s Next?
Overview
This article consolidates the known limitations of AI Agents and provides a troubleshooting guide to help you resolve common issues.
Behavioural Limitations
| Limitation | Description |
|---|---|
| Non-deterministic behaviour | AI agents may give slightly different responses to the same situation - this is expected for AI. |
| Context boundaries | Agents can only access and work with information within their assigned scope. |
| Language processing | Agents perform best with clear, descriptive text. Agents respect the language used in the instruction for their output. |
| Math/arithmetic unreliability | Agents are unreliable at arithmetic, counting, and multi-step calculations. Use formula fields and automations for calculation-heavy workflows. Agents are excellent at reading calculated values and deciding what to do based on them. |
| Similarly-named items | Agents can confuse siblings with similar names (e.g., "Gate1" vs "Gate2"). Split similarly-named items into separate actions. |
Actions AI Agents Cannot Perform
- Delete items
- Change task/project descriptions
Data AI Agents Cannot Access
- User workload metrics - agents can count assigned tasks but cannot access capacity planning or scheduling data.
- People-type custom fields
- Attachments (PDFs, images, documents)
- External data sources (knowledge bases, Google Docs, SharePoint, external APIs)
- Data outside the assigned scope of space
Consolidated Limits Reference
| Area | Limit |
|---|---|
| Subitems per execution (action target) | Max 1,000 subitems |
| Email recipients | Multiple "To" only - no CC/BCC, no attachments, no secondary/follower emails |
| Slack | One channel per action; no thread replies; no @mentioning |
| Microsoft Teams | Channels within a team only; one channel per action; no @mentioning; no thread messages |
| Location action | Can only move to folders referenced as location chips; cannot create folders |
| Agents per location | One agent of each type per folder/project/task |
| Date fields with time component | Not supported in agent actions |
Troubleshooting Decision Tree
If your AI Agent isn't responding:
- Is the agent toggled ON in Space Settings?
- Has the agent been appointed to a location? Creating an agent is not enough - you must @mention it in the target folder, project, or task.
- Does the trigger match what actually happened? For example, if the trigger is "Status changed" but the user changed a custom field, the agent won't fire.
- For date-based triggers, check that the date field has a value and the trigger timing is correct.
If your AI Agent is making wrong decisions:
- Check the Activity Dashboard and read the AI reasoning to understand how the agent reached its conclusion.
- Test in the Playground with the actual item that produced the wrong result.
- Make your instructions more specific - add explicit rules, examples, and fallback behavior.
- If the agent confuses similarly-named items (e.g., "Phase 1" vs "Phase 2"), split them into separate actions.
If assignment seems random or wrong:
- Make sure your prompt lists team members by name - agents cannot look up user groups or profile attributes in all cases.
- Check that names are spelled correctly.
- Include fallback logic for cases where no one matches your criteria.
- Remember that agents can count assigned tasks but cannot access workload or scheduling data.
What’s Next?
Please sign in to leave a comment.