Execution Advisor

  • Updated

Execution Advisor is the runtime intervention layer for specialized agents in Optimizely Opal. It works alongside Output Evaluation and Execution Guardrails to keep agent runs on course. Execution Advisor recovers runs that go wrong, retries failed tool calls, and reviews sensitive tool calls before they run.

Execution Advisor does not have its own configuration screen. Its interventions run automatically based on what Output Evaluation and Execution Guardrails detect. You see its effects in the agent execution trace and on the Execution Guardrails summary card.

Execution Advisor handles the following three scenarios:

  • Anomaly rescue – Recovers runs that Execution Guardrails flag.
  • Failure recovery – Recovers runs that hit terminating errors.
  • Sensitive tool approval – Reviews tool calls before they run.

Prerequisites

Anomaly rescue

Anomaly rescue saves runs that would otherwise be lost. When Execution Guardrails flag a run as deviating from normal behavior, Execution Advisor injects corrective guidance. This guidance steers the run toward a valid result.

Anomaly rescue runs in Watching and Enforcing modes only. During Learning mode, Execution Guardrails do not flag runs, so Execution Advisor has nothing to rescue.

When the rescue attempt succeeds, the run completes and counts as Rescued on the Execution Guardrails summary card. When the rescue attempt fails, the run's outcome depends on the guardrail mode. See Configure Execution Guardrails for details on outcome categories.

Failure recovery

Failure recovery keeps a run going when it hits an error that would normally stop the execution. Execution Advisor generates recovery guidance and attempts to recover the run with that guidance.

Failure recovery handles the following error types:

  • Tool call failures – API errors, timeouts, and invalid responses.
  • Runtime errors – Exceptions and stack traces from code the agent ran.
  • Malformed inputs – Inputs the receiving tool cannot process.
  • Other terminating errors – Any error that would otherwise end the execution.

Execution Advisor decides how to recover based on the following:

  • The error type and message.
  • The stack trace, if one is available.
  • The tool name and parameters involved in the failure.
  • The full conversation context up to the point of failure.
  • The agent's purpose and description.

The default is one rescue attempt per execution. You cannot configure the retry behavior.

When the rescue attempt succeeds, the run continues toward a valid result. The run counts as Rescued on the Execution Guardrails summary card. When the rescue attempt fails, the run terminates.

Sensitive tool approval

Sensitive tool approval prevents agents from performing destructive actions without a safety check. Before a sensitive tool runs, Execution Advisor reviews the call. Sensitive tools include those that send email, publish content, or delete data. Each reviewed tool call shows an inline verdict in the execution trace.

The verdicts are

  • Approved – The tool call meets safety criteria. Execution Advisor lets it run.
  • Blocked – The tool call does not meet safety criteria. Execution Advisor stops it.

When Execution Advisor blocks a tool call, the run counts as Unapproved on the Execution Guardrails summary card.

Optimizely maintains the list of sensitive tools. Agent builders cannot add or remove tools from this list.

Where to see Execution Advisor in action

Execution Advisor does not have its own user interface. You see its effects in two places.

Agent execution trace

The execution trace shows every step in a single agent run, including any Execution Advisor interventions. Open the trace from the Logs tab of a specialized agent.

  1. Go to home.optimizely.com.
  2. Select your organization.

  3. Click Opal.

    Screenshot of the Optimizely home page showing the Opal option in the navigation

  4. Click Agents.

    Screenshot of the Agents page in Opal where the list of agents is displayed

  5. Click the Your Agents tab.

    Screenshot of the Agents page in Opal where the Your Agents tab is selected
  1. Select a specialized agent.
  2. Click the Logs tab.
  3. Click a run row to open the execution details panel.

The execution trace shows the following Execution Advisor interventions:

  • Inline Approved or Blocked verdicts next to tool calls Execution Advisor reviews.
  • Retry attempts when failure recovery retries a failed tool call.
  • Corrective guidance Execution Advisor injects during anomaly rescue or failure recovery.

Execution Guardrails summary card

The Execution Guardrails summary card displays counts that reflect Execution Advisor's interventions.

  • Rescued – Runs Execution Advisor recovers through anomaly rescue or failure recovery.
  • Unapproved – Runs that include a tool call Execution Advisor blocks, or runs that Opal excludes from the baseline for other reasons.

See Configure Execution Guardrails for the full list of outcome categories.

Related articles

If you use Opti ID, administrators can turn off generative AI in the Opti ID Admin Center. See Turn generative AI off across Optimizely applications.