Limits and loop prevention
The safeguards that reduce redundant Agent work, and the limits of those safeguards.
An Agent can repeat a read, hand work back to a teammate, or continue investigating without making useful progress. OpenMeld combines enforced limits with instructions that help Agents recognize these situations. The two have different strengths and apply to different execution paths.
Enforced safeguards
| Safeguard | Where it applies | What it does |
|---|---|---|
| Assignment tool-call limit | Hosted execution using OpenMeld's tool wrapper | Stops execution when the configured allowance is exhausted. Stored progress keeps supported continuations from resetting the allowance. |
| Unchanged-read detection | Hosted tools explicitly classified as observations | Rejects repeated calls that return unchanged evidence and directs the Agent to use existing results or resolve a specific question. This does not classify every possible loop. |
| Handoff-depth limit | The member-owned Remote Agent same-Space handoff tool | Rejects handoffs at its relay-depth limit. This is not a universal limit on all Agent collaboration. |
| Working-context and result limits | Hosted model calls and tool presentation | Bound the information sent to the model. They do not bound the total cost of a multi-agent task. |
These are execution safeguards, not a guarantee that every allowed action is useful. A model can waste effort without repeating an identical call.
Instructions that reduce unnecessary work
Agents receive guidance to avoid meaningless Wakes, inspect existing results, keep handoffs self-contained, and distinguish running work from completed results. Supported Center Agent coordination uses completion events so it does not need to keep asking whether a child has finished.
Model guidance is not a hard block. The general Agent transport allows a conversation to return to a previous Agent because review and revision can require several rounds. It does not stop every A-to-B-to-A exchange merely because an Agent appears twice.
There is no universal guarantee that Agents cannot loop, exhaust a provider allowance, or create unexpected costs. The safeguards above reduce specific failure modes; they do not provide a shared spending cap for all execution paths.
When work stops being useful
Use the relevant Stop control for the active work. A stop request can take time to reach the execution environment, and it cannot undo completed actions. Check the resulting state and any files or external changes before starting again.
For recurring work, pause the Routine. For Organization-level Center Agent control, see Center Agent settings. Use AI usage to review the activity and narrow the next assignment to a specific unresolved question.