Compensating Action

In banking or e-commerce, a compensating action semantically undoes a completed financial operation.

• Example: An autonomous agent executes a funds transfer of $100 from Account A to Account B. A subsequent error requires compensation.
• Compensating Action: Initiate a new, separate transfer of $100 from Account B back to Account A. This is a business-level undo, not a database rollback.
• Key Consideration: The system must handle potential side effects like fee calculations or ledger entries triggered by the original transaction.

In supply chain systems, reserving stock is a committed action. Compensation releases that hold.

• Example: An agent reserves 50 units of a product for a customer order. The payment subsequently fails.
• Compensating Action: Release the 50-unit reservation, incrementing available inventory. This may involve notifying other agents that the stock is now free for other orders.
• Complexity: The compensation must account for temporal changes, like price updates or partial fulfillments that occurred after the initial reservation.

When an agent calls an external service that changes remote state, compensation often requires a separate API call.

• Example: An agent calls a cloud provider's API to launch a virtual machine. Later steps fail, requiring cleanup.
• Compensating Action: Call the cloud provider's termination API for that specific VM instance. The agent must store the instance ID from the original call to execute this.
• Idempotency is Critical: The termination call must be safe to retry in case of network failures during the compensation phase.

The Saga pattern is a distributed transaction model built around compensating actions. Each step in a long-running workflow has a defined compensating action.

• Process: A saga coordinates a sequence of local transactions (T1, T2, T3...). If T3 fails, the saga executes the compensating actions for T2 (C2) and T1 (C1) in reverse order.
• Real-World Use: Travel booking: booking a flight (T1), hotel (T2), and car (T3). If the car booking fails, compensate by canceling the hotel (C2) and flight (C1).
• Design Principle: Compensating actions must be idempotent and commutative to handle retries and out-of-order failures safely.

In content management or legal systems, publishing a document is an action. Compensation marks it as superseded or withdrawn.

• Example: An autonomous legal agent files a generated motion with a court docket system. An error is found in the reasoning.
• Compensating Action: File a formal "Notice of Errata" or "Withdrawal of Motion" document. This updates the docket's state, informing human actors of the change.
• Semantic Undo: The original filing remains in the audit log, but the system's active state reflects the new, corrective status.

Sending an alert is an action with external impact. Compensation involves sending a corrective follow-up.

• Example: A monitoring agent sends a critical outage alert to an engineering team. The agent's own diagnostics later determine the alert was a false positive.
• Compensating Action: Immediately send a follow-up "All Clear" or "False Alarm" notification through the same channels (e.g., PagerDuty, Slack, email).
• Goal: Mitigate the operational cost (context switching, investigation time) caused by the erroneous initial action.

Action rollback is the process of reverting the technical effects of a specific executed action to restore a system to a previous, consistent state. It is often a lower-level, mechanistic reversal compared to the business-logic-aware nature of a compensating action.

• Scope: Typically applies to a single operation or database transaction (e.g., ROLLBACK in SQL).
• Contrast: Rollback is a technical undo (state reversion); a compensating action is a semantic undo (logical counteraction).

Plan repair is the process of modifying a partially executed or failed plan to achieve the original goal. It involves analyzing the failure, the current state, and the goal to generate a revised sequence. A compensating action may be one step within a larger repair strategy.

• Techniques: Includes action substitution, step reordering, or constraint relaxation.
• Relationship: After a failed action, an agent may execute a compensating action and then invoke plan repair to find a new viable path forward.

Fallback execution is a fault-tolerant strategy where a system switches to a predefined alternative action or workflow when a primary operation fails. While a compensating action undoes a previous step, a fallback action provides an alternative way to achieve a similar outcome.

• Resilience Pattern: A core component of circuit breaker and retry logic implementations.
• Sequencing: A system might first retry, then execute a fallback, and if that fails, finally trigger a compensating action for preceding steps.