> For the complete documentation index, see [llms.txt](https://docs.tread.fi/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.tread.fi/agents/agent-analytics.md).

# Agent analytics

The Agent analytics page (Agents → More → Agent analytics) shows what your agents did over a period: the orders they placed, how those orders ended, how well they executed, and what they cost in LLM tokens. This page explains every panel and the data behind it. The page itself carries only the figures; the "Analyze with Execution Quant" button opens a chat with an agent that can explain any panel in plain words and show it beside its answer.

### Scope and filters

* **Range** is a rolling window that ends now: 1D, 7D, 30D, 90D or 1Y. There is no all-time view.
* **Exchange** and **pair** narrow the order panels and the market-making panels. Token usage, Agents used and Agents the supervisor calls ignore them.
* **All (incl. deleted agents)** adds agents you can no longer see (deleted, or a shared role that was revoked). By default they are left out everywhere, including the manual benchmark.
* All times are UTC. On 1D the Order activity axis shows your local time.
* Only orders your agents placed are counted. An order is counted once per order row, so a retried placement counts twice; Outcome breakdown de-duplicates. A take-profit or stop-loss leg that later closed an entry is part of that entry, not a second order, though the volume it executed still counts.
* **Manual** means orders no agent placed. Every comparison on the page uses that one definition. An agent you can no longer see is on neither side. A scheduled basket's or a hedge's legs belong to the agent that placed the parent.
* Token usage and Agents used come from daily records, so they cover whole UTC calendar days inside the range (7D is the seven UTC dates ending today), not the rolling hours the order panels use.
* No PnL or fees appear on the Overview, Execution or Resources tabs. Only the Live snapshot shows PnL, and that is gross (mark-to-market, fees excluded).

If some trading data could not be loaded, the page says so above the panels and offers a retry. Panels built on the missing read are then empty because of the outage, not because nothing happened, and the execution score is left out for that load.

### Execution score

The score is an absolute 0–100 for every order-placing agent: market-making, desk and delta-neutral alike. It measures how far each order's fills landed from the order's own reference price, in units of that market's normal 15-minute move, through one fixed curve.

* **50** means filled at the reference price. Roughly **70** is half a normal move better than the reference and **30** half a move worse. It is not a rank, a percentile or a position in your own book, and it has no fee or fill-speed terms.
* **Bands:** Excellent ≥80 · Good 60–80 · Average 40–60 · Poor <40.
* **Units.** One scored unit is one desk order, one settled market-making pair, or one delta-neutral hedge (both legs together).
  * Desk scheduled orders (TWAP, VWAP, IS, Target Time): cost = ½ arrival slippage + ½ VWAP slippage, scaled to the order's own horizon, so a long TWAP is judged against a bigger move than a quick one.
  * Market making: spread captured per dollar traded (gross PnL over volume, fees excluded). This is a cost measure, not a PnL ranking.
  * Delta-neutral: the notional-weighted cost of both legs, on a tighter scale than the other two.
* **•** next to a score means it rests on fewer than 30 scored units. Treat it as provisional.
* **vs manual** is the plain difference between the agents' mean and your manual orders' mean, both scored the same way. There is no significance test. A mean depends on the mix of order types, because market-making, desk and delta-neutral units do not share a median, so compare market making with market making and desk with desk before reading a gap as execution quality.

### Overview tab

**KPI cards.** Executed volume is the USD notional your agents executed in the window, from fills, with the order count under it. Orders counts the agent orders placed, with the number of agents that placed any. LLM tokens is every agent's tokens in the window, read-only agents included, with the turn count. Execution score is the agents' score with your manual benchmark under it and a "pts vs manual" badge; it reads — when fewer than 10 agent orders could be scored and n/a when none has a benchmark.

**Order activity.** Orders per UTC day (per hour on 1D), stacked by agent. Only days with orders are drawn. Hover a bar for each agent's count and the total.

**Agents used** (the toggle on the same card). One bar per UTC day, even on 1D. Bar height is the number of agents that ran that day, counting any turn or tokens, so read-only agents appear here and not in the order view. Hover a bar for each agent's tokens.

**At a glance.** Up to three lines: the agent with the most executed volume and its share of all agents' volume; the agent with the best execution score and its points against the manual benchmark (score alone when there is no benchmark); and the agent that used the most tokens and its share. A line is left out when nobody has that figure.

**Agent performance.** One row per agent that can place orders and traded in the window. Read-only agents and the Supervisor, which trades through child agents, are not listed. Columns: order share, orders, completion (COMPLETED out of all orders), executed volume, tokens and execution score. Hover a score for how many units it rests on and why a row has none.

**Supervised vs direct.** An order is supervised when the agent job that placed it runs under a supervisor; this uses the job's final state, so a job adopted later counts as supervised for all of its orders. The bar is the supervised share of orders. Orders, Completion, Executed volume, Execution score and the Order mix cover every order on each side. MM pairs and MM volume cover settled market-making pairs only, and a side with none shows —.

**Outcome breakdown.** How each agent's orders ended, each order counted once. A market-making parent that has settled uses its warehouse outcome: Completed, Hit TP or Hit SL. A pair cancelled after some fills is *Cancelled by you* when the parent was plainly cancelled, and *Cancelled by system* for a stop-loss or take-profit cancel, a failure, or a duration that ended with leftovers. Every other order uses its status: COMPLETED is Completed, CANCELLED is Cancelled by you, FAILED or EXPIRED is Cancelled by system; anything else is the grey segment with no legend entry. A row says "parent orders" only when every one of that agent's orders is a multi-order.

**Token usage.** A donut of the window's tokens with each agent's share, read-only agents included. *Weekly budget used* shows tokens used against the account's weekly cap for the current UTC week (Monday 00:00 to the next Monday). It is account-wide, there is no per-agent budget, and it ignores the range and filters. The bar turns red at 90% and stops at full; the percentage can pass 100% when the account is over its cap.

### Execution tab

**Execution quality.** The execution score split into All agents, Your manual orders (when you have any) and, when more than one agent has units, each agent. Each row shows its band mix, its score and how many units were scored. Chips under the rows split the agents' units by kind: Market making, Desk, Delta-neutral. Hover a score for its sample size, confidence and, for an unscored row, the reason.

**SL / TP trigger rates.** Two grids, Stop-loss hit rate and Take-profit hit rate, by distance (percent, rounded) and leverage bucket. Each cell is hits out of pairs. The toggle switches between your agents' pairs and your whole book. The stop-loss grid counts only pairs that had a stop-loss set; pairs with no distance or no leverage bucket are left out.

**Pairs traded.** The share of agent orders per pair, top eight plus Other, across every order type. Orders with no pair are left out.

**Activity heatmap.** Agent orders by UTC weekday and hour. Darker cells mean more orders; hover a cell for the count.

**Your agent runs vs your manual runs.** Desk orders only, and shown only when both your agents and you placed desk orders in scope. Arrival cost and VWAP cost are in basis points, weighted by notional; positive means cost, and the better side of each row is highlighted. When some pair-and-exchange combinations were traded by both sides, only those are compared and the notional row reads *Matched notional*; that removes the pair and venue mix from the comparison. Without any overlap the split is unnormalized, so treat differences as mix-driven.

### Resources tab

**Token usage.** The same panel as on Overview.

**Orders placed by agents.** A donut of the window's orders per agent, with agents under a 5% share folded into Other, and the share of orders that ran under a supervisor. When some did, *Supervised orders by agent* follows: for each agent, the bar and the Supervised column are that agent's own supervised rate, and Share is its slice of all supervised orders.

**Agents the supervisor calls.** Child agent jobs the supervisor started in the window, one per delegation, most to least. It counts delegations, not orders, so an agent that was called but never traded still appears. It ignores the exchange and pair filters.

### Live snapshot tab

The Live snapshot streams from your running tasks right now. It ignores the range, exchange and pair filters, and nothing on it is historical. The header counts the running top-level tasks.

**Tiles.** Volume is the running tasks' volume added up; a supervisor task includes its child agents. Margin used is the margin behind the open positions, from account balances, with leverage and health; it reads *not priced* when the balances feed cannot price it. Net exposure is long minus short across open positions. Gross PnL is mark-to-market and excludes fees; the line under it splits the fleet's books into realized (the FIFO-matched result on closed inventory) and unrealized (open inventory at the current mark). The two are computed book by book and need not add up to the headline, because delta-neutral books carry funding that only the headline includes.

**By agent.** One row per agent, adding up that agent's running tasks. Agents a supervisor is running sit under the Supervisor row; expand it to see them, and expand any agent to see its tasks and open a task's chat. Columns are Tasks, Orders, Volume and Gross PnL (fees excluded). Delta-neutral agents show dashes for orders, volume and PnL because their live figures carry exposure instead.

**Venues traded.** Where the fleet's open book sits, broken down into the accounts behind each venue. Share is the venue's share of exposure. Pairs counts open legs, not distinct pairs. Leverage and margin health are per account and dash wherever the balances feed has not priced one.

**Recent activity.** The newest orders and placement outcomes across your running agents, newest first, with the agent, venue and age of each. A sliced order's partial fills appear as one row per fill. Market-making and delta-neutral rows show — for account, because their fills carry no account.
