> ## Documentation Index > Fetch the complete documentation index at: https://docs.tied.com/llms.txt > Use this file to discover all available pages before exploring further. # Operations > Watch it work, audit what it did, step in, and correct it systematically. Four things a support manager needs and most vendors hide. Running an AI agent in production is a management job, not a configuration job. Four capabilities make it possible. A live trace of the turn as it executes. Any transcript, with the reasoning and the action receipts attached. One-click takeover of a live conversation. A queue of the resolutions most likely to be wrong. ## The live trace Open a conversation the agent is handling and watch the turn execute: the router picking an intent, a specialist running, retrieval returning, a tool being called, the composer writing. **The panel names real components.** The router, the specialists and the composer are genuinely separate things with their own tools and their own tests — so what the panel shows is the actual call graph, not a progress animation with labels on it. When something goes wrong, the trace is where it went wrong. That matters most on the day you have to explain the agent's behaviour to *your* customer. A window into execution can be quoted. An animation cannot. ## The reasoning trace, on purpose Most vendors hide this. The stated reason is complexity or intellectual property; the practical effect is that you cannot tell why the agent did what it did, and you cannot argue with an invoice built on top of it. Tied shows it — per turn and per resolution: * what was retrieved, and how strongly it scored; * which specialist ran and what it returned; * every action, with its arguments, its authorization, and the provider's own transaction ID; * why the composer said what it said; * if it escalated, **which of the seven grounds fired**, and the evidence that made it fire. Showing this is only safe because it comes with the receipt. ## The receipt Every billed resolution carries one, and it is the enforcement mechanism for everything Tied claims about [pricing](/pricing/resolutions). | The receipt shows | Which lets you check | | - | - | | The **transcript reference** | Exactly which conversation this charge came from | | Which **signal tier** confirmed it, and **which specific corroborator** | Not "verified" — *which* link was opened, *which* system state changed, *which* login succeeded | | The **class and its discriminator** | The specific trail entry that fixed the price | | The **orchestration fee** | The fee is the fee you were quoted | | The **itemised model cost** — model, tokens, rate, total | That the markup really is zero | | A **one-click dispute** | That disagreeing is cheap | **One click.** Not an email address, not a support queue, not a form that goes to a team that will get back to you. If a receipt does not hold up, saying so should take about as long as noticing. ## Taking over One click moves a live conversation from the agent to you. The agent stops, the context is already there, and you carry on. **A human message before a resolution is confirmed means no fee.** Not discounted — zero. Once a human sends a message in the thread, the outcome came from the human. Once a resolution is confirmed, its fee is locked, and a later message or takeover does not cancel it. See [What counts as a resolution](/pricing/resolutions). That is the incentive Tied deliberately chose. A vendor that could bill on handoff is a vendor paid for giving up, and every incentive quietly bends toward escalating early and often. ## The quality queue Tied credits you automatically on a negative signal, which means a bad resolution costs us money. The quality queue is how that exposure stays bounded — and it is the most useful hour of your week. It collects the cases most likely to be wrong: Ambiguous signal. Billing already resolved them in your favour; the queue is where you find out whether the agent was actually right. Already voided and credited automatically. Still worth reading, because the credit fixes the invoice and not the cause. Reopens inside the 30-day window, and anything else that voided. Working an item produces one of three outcomes: the agent was right, the agent was wrong and here is the correct answer, or the knowledge was missing and here it is. **Every outcome routes into an [evaluation golden set](/agents/evaluation).** That is what turns the queue from an apology desk into a data source — a real failure becomes a permanent regression test rather than a one-off fix. ## Reporting The performance view reports **confirmed resolutions only**, with escalation and abandonment accounted for explicitly alongside rather than dropped from a denominator. Because reporting and billing share one definition, the numbers reconcile against your invoice line for line. A resolution rate that cannot be reconciled is worse than no dashboard — it becomes the thing you dispute instead of the thing that builds trust. ## Next Where queue outcomes end up, and what gates a release. The signal tiers, the void windows, and the guarantee. What your team sees when the agent hands off. Account-level record of configuration changes. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.