Platform9 min read

Operation Center

The single place to review agent work and monitor activity across your agents and Personal Assistant.

Overview

The Operation Center is the hub for everything your agents and Personal Assistant do. Open it from the Operations item in the top navigation.

It has two tabs:

  • Review Queue — approve, edit, override, or decline pending actions
  • Activity — a live feed of everything your agents have done

Review Queue

Every time an agent or your Personal Assistant proposes an action — sending a reply, changing a ticket status, creating a Linear issue, writing an internal note — the result passes through the review system. Depending on how each agent is configured, the action is either delivered automatically or held for a human to review.

Each review moves through a simple lifecycle:

  • Pending — waiting for a human to review
  • Approved — the agent's proposed actions are sent unchanged
  • Edited — a reviewer adjusted the content of one or more actions before sending
  • Overridden — a reviewer changed which actions run (added, removed, or replaced an action)
  • Declined — rejected; no action is taken
  • Dismissed — a Needs attention item that a reviewer cleared without adding an action

Needs attention is a badge on a pending item rather than a status of its own: the agent was triggered but produced nothing — no reply, no action, no proposal. Two run warnings produce it — No output produced from the uploaded file (a file came in that the model could not read) and Agent took no action (the agent ran but never responded). If something was needed, add the missing action and submit it; if nothing was, use Dismiss. Dismissing is final — the item becomes Dismissed and cannot be reopened. There is no Needs-attention or Dismissed filter tab; dismissed items are found under Handled or All.

Needs-attention items are queued no matter how the agent's review mode is set, including Autonomous — a run that produced nothing is always worth a human's eyes.

Filtering and Scope

Use the filters at the top of the queue to narrow what you see:

  • Filter by agent to focus on a single agent or your Personal Assistant
  • Filter by status to focus on pending, approved, edited, overridden, or declined items
  • Filter by date range to review a specific time period

If you have personal reviews enabled, a Scope toggle lets you switch between agent reviews, Personal Assistant reviews, or both at once. A badge on the Review Queue tab shows how many Personal Assistant items are waiting for you.

Reviewing an Item

For a quick decision, each queue row carries its own controls — icon buttons on desktop, labelled buttons on mobile. Approve sends all of that row's proposed actions at once and only appears when there are proposals to send; Decline and, on needs-attention rows, Dismiss are available in their place. To approve some of a row's actions but not others, open the detail view.

Click any item to open its detail view. You'll see:

  • The original context (ticket, email, chat, or scheduled trigger that started the work)
  • The agent's proposed actions, grouped by type — some also show who the message would go to (see Who the message is addressed to below)
  • A trace of the tools the agent called to produce the result
  • Any actions the agent has already completed automatically, shown for context
  • Approve as-is, Edit & approve, Override, and Decline controls — untouched proposals are kept and approved as-is by default; Edit & approve changes an action's content, while Override changes which actions run (add, remove, or replace one). You can also attach a note to record why you made a change. On an item marked Needs attention there is nothing to approve, so the controls change: under Or add an action the agent didn't propose, pick the action the agent should have taken and Submit it — or, if the run genuinely needed no follow-up, choose Dismiss.

Who the Message Is Addressed To

Where Symbi can work out the destination of an outbound message, the detail view shows it above the draft: the address, the signal it came from (last inbound reply, contact of record, or sender of the message being answered), and how long ago it was worked out. The provenance is the part worth reading — it is how you catch a colleague who once wrote into the thread becoming the destination of a supplier request.

The address is a snapshot, not a commitment. It is resolved once, when the agent runs, and resolved again when the message is actually sent — so on a proposal left overnight, check the "resolved n hours ago" stamp before you approve.

If the destination could not be worked out, one of two warnings appears in its place, and the difference matters:

  • No recipient. (red) — there is genuinely nobody to write to, and the platform refuses the approval. Override to retarget the action, Decline it, or re-run the agent once there is someone to answer.
  • Recipient unknown. (amber) — the integration did not answer in time. Approving is still allowed: the address is worked out again at send time, and the send fails loudly if there is really nobody there. Don't decline on this alone.

Warning

Not every outbound action carries this line. It is resolved per tool by the app the message goes through, which today means Zoho Desk replies and Gmail replies and drafts. For every other action no line appears at all, and that absence says nothing about whether the recipient is right — read the draft itself.

Multi-Action Reviews

A single task often produces multiple actions. For example, handling a support ticket might combine:

  • Responding to the customer
  • Changing the ticket status (e.g., from Open to Pending)
  • Adding an internal note for the support team

Each action is reviewable independently. You can approve the customer reply while declining the status change, or edit the internal note without touching the others.

Actions are grouped into five tool categories to keep the review clear:

  • Communicate — anything the agent sends to a person: customer replies and emails, Slack messages and DMs, calendar invitations, and comments on issues or pull requests
  • Status & Lifecycle — ticket or record status changes
  • Organize & Categorize — categorization, tagging, and prioritization
  • Internal Notes — internal notes and comments
  • Produce Output — data writes to spreadsheets, registers, and external records

Bulk Actions

For high-volume teams, select multiple reviews with the checkboxes and choose Approve or Decline from the toolbar to clear the queue in one step. Needs-attention rows have no checkbox — with no proposals to approve, a bulk approve would record them as declines — so handle those one at a time in the detail view, either by adding an action and submitting it or by dismissing.

Slack Reviews

If Slack is connected, reviews can also be approved, edited, overridden, or declined from an interactive message in your Slack channel. Actions taken in Slack update the Operation Center instantly, and vice versa — your team can work from wherever is most convenient. Needs-attention items reach Slack only when the agent itself flagged the run; that message offers only Dismiss and Decline, since nothing was proposed to approve, edit, or override. Items the platform raises on its own — for example a run that produced no output — appear only in the Operation Center, so check there too. See Connecting Slack for setup.

The Slack message carries the same recipient line as the detail view, on every proposed message that has one. It also enforces the same refusal: when a message has no recipient, Approve as-is and Edit & approve are left off the message entirely, and only Override… and Decline remain. Buttons missing from a review message mean the platform will not accept an approval, not that Slack failed to render.

Activity

The Activity tab is a live feed of every task your agents have processed, whether it was auto-approved or reviewed. Use it to:

  • See what agents are doing right now
  • Spot patterns in recent work
  • Drill into a specific task to understand how an agent reached its conclusion

Activity includes both agent tasks and Personal Assistant runs (if you have personal permissions enabled).

Review Modes

Each agent can be configured with one of three review modes that decide whether its actions go to the Review Queue or are delivered automatically:

  • Always Review — every action is queued for human review before delivery
  • Confidence-based (X%) — the AI evaluates its own confidence. Actions above the threshold are auto-approved; lower-confidence actions are queued for review
  • Autonomous — every action is delivered automatically

Tip

Start new agents in Always Review. Once you're comfortable with how they handle real work, move to Confidence-based and raise the threshold gradually. Switch to Autonomous only for tasks where mistakes are cheap.

Agent-Mode: Per-Tool-Group Review

Agent-mode agents let you set a different review mode for each of the five action groups independently:

  • Communicate — anything the agent sends to a person: customer replies and emails, Slack messages and DMs, calendar invitations, and comments on issues or pull requests
  • Status & Lifecycle — ticket and record status changes
  • Organize & Categorize — categorization, tagging, and prioritization
  • Internal Notes — internal notes and comments
  • Produce Output — data writes to spreadsheets, registers, and external records

For example, you can set Communicate to Always Review while Organize and Internal run Autonomous — keeping approval gates on outbound replies while the agent handles internal work without interruption. Configure per-group review from the What it does section on the agent's page. Changes take effect when you publish the agent.

These modes govern what the agent sends. Symbi's own review notifications in Slack are platform delivery, not agent actions, and are never held for approval by the Communicate setting.

Auto-Extending Your FAQ From Edits

When you edit an action before approving it, your correction is captured. If Auto-extend FAQ from support tickets is enabled for that agent, a daily distillation run analyzes those approved and edited reviews, extracts FAQ patterns, and adds them to the agent's knowledge base — so your corrections accumulate into lasting quality improvements rather than being applied only once.

Info

The more you review and refine, the faster your agents improve. Most teams need fewer edits within a few weeks of consistent reviewing.

What's Next?