Agents

Agent scope

Agents should start with a narrow job and a clear audience. Define what the agent can answer, which tools it can use, what it must refuse, and who reviews traces. Do not give an agent broad write access before evals and traces show reliable behavior.

Tools and memory

Agent quality depends on selected Knowledge Base sources, tool permissions, triggers, memory policy, and review surfaces. The Agents feature explains the builder, and roles and permissions should be reviewed before granting broader capability.

Evaluation

Use low-risk tests before publishing an agent to a customer-facing panel. The agent tutorial shows how to keep checkout knowledge answers bounded and reviewable.

Quick reference

• Use this page when turning repeatable WordPress, WooCommerce, marketing, support, reporting, agent, or tool work into a controlled process.
• Do not enable unattended write behavior, external calls, or customer-facing output before failure handling and approval ownership are clear.
• Key decision: whether the capability has narrow permissions, negative tests, useful audit records, and a safe unavailable-state behavior.

Scope limits

• Do not use this page to approve broad tools, unsafe external calls, customer-facing agents, or visitor panels without negative tests and audit evidence.
• This page does not justify unattended execution when trigger scope, owner response, failure handling, or approval behavior is unclear.
• Use it to narrow repeatable work before expanding to write actions, external services, or customer-facing output.

Owner and cadence

• Primary owner: operations lead for the affected workflow, watcher, agent, playbook, or custom tool.
• Review cadence: before first run, after failed runs, after provider changes, and during monthly workflow review.
• Escalate when workflow execution writes production data, repeats failures, sends customer-facing output, or runs without a visible owner.

Access and data boundary

• Agent, app, tool, and webhook evidence should name allowed sources, tools, payload fields, permissions, external destinations, negative tests, and audit fields.
• Grant workflow, watcher, playbook, agent, and custom-tool access by role, record type, launch surface, and risk level.
• Start with read-only, staging, simulation, or notification-only runs before allowing production writes, external webhooks, or unattended execution.

Production checklist

• Give each agent a narrow purpose, approved sources, explicit tool permissions, refusal rules, trace review, and staged rollout.
• Run eval cases before exposing an agent to customer-facing panels or write-capable tools.
• Define trigger, owner, input data, output, approval requirement, retry behavior, failure notification, and kill switch before enabling a workflow.
• Start with read-only runs or staging examples until the team has reviewed successful traces and audit records.

Acceptance checks

• Agent traces show bounded behavior and explain why tools or sources were used.
• A reviewer can identify when the agent should refuse, escalate, or request human approval.
• The workflow or agent has a named owner who can pause it and explain its last run.
• Failures produce enough audit, diagnostics, and notification context for another operator to respond.

Failure modes to test

• Test failed evals, unsafe tools, malformed payloads, permission denial, external-service errors, replay attempts, missing audit fields, and visitor fallback states.
• Test missing trigger data, duplicate runs, retry exhaustion, failed notifications, stale sources, permission denial, tool errors, and provider limits.
• Confirm write actions remain paused or approval-gated when simulation, staging, or first-run evidence fails.

Evidence to capture

• Capture agent scope, sources, tools, eval set, failed cases, trace review, launch surface, and approval owner.
• Record trigger, owner, input scope, affected records, approval point, run ID or timestamp, output artifact, and failure response.
• Capture simulation, staging, notification-only, or first production run evidence before expanding workflow scope.

Decision record

• Decision field to include: allowed sources, allowed tools, permission scope, negative-test result, trace or payload evidence, and launch surface.
• Record the workflow decision, trigger, owner, input scope, output artifact, approval point, retry rule, failure notification, and kill switch.
• Include the first-run evidence, run ID or timestamp, cost expectation, expansion criteria, and the condition that keeps write actions paused.

Stop or rollback path

Unpublish or narrow the agent when evals fail, traces are unclear, tools are too broad, or customer-facing answers cannot be explained. Relaunch after source, tool, and refusal-rule changes pass review.

Monitoring window

• Review traces, eval failures, tool-call errors, refusals, citations, and memory behavior after real requests.
• Monitor first runs, retries, costs, failures, alert volume, queued work, and owner response time.
• Review run history before enabling write actions, more triggers, or broader records.

Expansion criteria

• Agents, tools, or webhooks can expand only after evals, schema failures, permission checks, audit fields, and negative tests pass.
• Expand workflow execution only after first-run history shows expected output, manageable alert volume, known cost, and clear failure handling.
• Write-capable steps remain approval-gated until staging, simulation, or notification-only runs are stable.

Common mistakes

• Turning a useful prompt into automation before defining trigger, owner, input scope, approval rule, and failure handling.
• Ignoring noisy alerts or failed runs until operators stop trusting the workflow surface.

Common questions

What should block agent, app, tool, or webhook launch?

Block launch when permissions are broad, schema is vague, evals fail, traces are unclear, external retries are unsafe, or audit records cannot explain the behavior.

Who should own this decision?

The operations owner should own the workflow, with specialist reviewers for customer, commerce, content, support, or tool impact.

What should stop the rollout?

Stop when trigger scope, owner, run state, approval behavior, failure response, or write capability is unclear.

Related operations

• Use Workflow Safety before enabling recurring workflows.
• Use Workflow Safe Mode and Kill Switches before production workflow rollout.
• Review Audit Log Review after the first production runs.
• Use Model Evaluation and Regression Review before broad agent or workflow rollout.
• Use Playbooks and Quick Actions for repeatable structured tasks.
• Use Prompt Template Governance before sharing reusable instructions.
• Use Playbook Import Export and Agency Reuse before reusing client workflows.
• Use Tool Validation and Schema Testing before exposing custom tools.