AI Agent Tool Retries: How to Handle Failed Function Calls Without Duplicates

Agent workflows depend on external services. Networks fail, APIs time out, and providers return temporary errors.

Retries are useful, but retries around side effects can be dangerous.

Imagine an agent calls create_invoice. The provider creates the invoice but the response times out. The agent sees an error and calls create_invoice again.

Now two invoices exist.

Separate retryable and non-retryable errors

A timeout may be retryable. An authorization failure usually is not.

A malformed argument should not be retried unchanged.

A rate limit may be retried after a provider-specified delay.

Make this distinction explicit in the tool layer rather than asking the model to decide blindly.

Use idempotency

Important side-effecting tools should accept an idempotency key.

For example:

agent_workflow_123:invoice:customer_456

If the same operation arrives again, the service can return the original result instead of creating another resource.

Bound retries

Retries need limits.

Set maximum attempts, exponential backoff, total workflow time, and a final failure state.

Without limits, a model can turn a temporary dependency failure into a long-running loop.

Tell the model what happened

A timeout does not necessarily mean the operation failed.

Return a structured status such as:

unknown_execution_state

That should cause the workflow to check the resource before creating another side effect.

Do not retry human approvals

An approval step is different from a transient API failure. Do not automatically replay an approval request because a downstream operation timed out.

Preserve the original approval context and ask whether the specific operation completed.

Final takeaway

Retries should be designed in the trusted tool layer.

Classify errors, use idempotency keys, cap attempts, distinguish unknown execution from confirmed failure, and never let the model blindly repeat a side effect.

Source: distributed-systems reliability and API idempotency principles.

Production implementation

The tool execution layer should remain responsible for authentication, authorization, validation, rate limits, retries, timeouts, and logging. The model should receive a narrow interface and a sanitized result rather than direct access to infrastructure. This separation lets engineers change providers without changing the agent's conceptual capability.

Before shipping a tool, test normal inputs, malformed arguments, unauthorized resources, provider failures, duplicate requests, slow responses, and cancellation. For side-effecting tools, verify that retries cannot create unintended duplicates. For read tools, verify that returned data contains only what the agent needs.

Keep the tool contract versioned when it becomes important to production workflows. Changes to argument names, required fields, enum values, or output shape can affect prompts and orchestration logic. Treat those changes like API changes rather than casual prompt edits.

Final takeaway

Reliable tool calling is mostly good software engineering around an unreliable model. Give the model clear capabilities, then put deterministic controls around execution. The result is an agent that can recover from ordinary failures without turning every failure into another guess.

Production design

A production tool layer should be deliberately boring. The model chooses from a documented capability set, while deterministic application code handles the parts that must not depend on model behavior. Validate the request, authenticate the caller, check authorization, validate resource ownership, apply rate and cost limits, execute with a deadline, and return a sanitized result.

Keep tool execution separate from the prompt layer. This makes it possible to change the model without changing the security boundary. It also gives engineers one place to add logging, metrics, retries, circuit breakers, and provider-specific behavior.

Failure scenarios

Test more than a successful request. Simulate malformed arguments, missing fields, unauthorized resources, expired sessions, rate limits, provider outages, slow responses, duplicate requests, partial responses, and cancellation. For side effects, deliberately create the condition where the provider succeeds but the response is lost. The application should recover without creating a second side effect.

For parallel workflows, test dependency races and partial completion. If three tools run concurrently and one fails, define whether the workflow can continue, whether completed operations should be compensated, and what the model should be told.

Observability

Record a workflow ID and tool-call ID for every execution. Useful fields include agent identity, tool version, operation, resource, authorization result, latency, retry count, error category, and final outcome. Do not place secrets or unnecessary customer data into the model-facing result or logs.

Metrics should distinguish model errors from tool errors. A spike in invalid arguments suggests a schema or prompt problem. A spike in timeouts suggests an infrastructure problem. A spike in permission denials may indicate a product workflow problem or an attempted abuse pattern.

Versioning and rollout

Treat important tool schemas like APIs. When a required argument changes, version the contract or provide a compatibility layer. Roll out significant changes gradually and monitor error rates before removing the previous version.

Keep production and development tools separate. A test agent should not accidentally discover a production endpoint merely because the tool registry is shared.

Security boundary

The model is an untrusted decision-maker. The tool executor is the trusted enforcement layer. This distinction should remain true even if the model is highly capable. Prompts can explain policy, but only application code should enforce permissions, resource scope, validation, and side-effect controls.

Final takeaway

Reliable tool calling comes from combining a clear interface with deterministic execution controls. Give the model enough capability to complete the task, but keep the actual authority in code that can validate, limit, observe, and stop every operation.

Testing checklist

Before production, create automated cases for valid and invalid schemas, missing required fields, unknown enum values, oversized inputs, unauthorized resources, provider failures, repeated requests, and slow dependencies. Verify that the tool returns a stable error category and that the workflow does not accidentally retry a non-retryable failure.

For side-effecting operations, run the same request twice and verify the business result remains correct. For concurrent operations, verify that independent calls can run together while dependent calls preserve their required order. For cancellation, confirm that an aborted workflow does not leave an uncontrolled background operation running.

These tests should run in CI because tool contracts change as the product evolves. A tool is part of the agent's public behavior even when it is technically an internal function.