Tracer
Live observability for deployed brokers — see the network, trace tasks, and read the LLM's reasoning.
Tracer answers one question about a broker that's already live: "someone talked to my agent — what did it actually do, and why did it decide that?" It's read-only observability for deployed Agent Networks.
What Tracer does
You do four things in Tracer:
- Pick a scope — business group → environment → deployed broker → time window.
- Read the topology — a live graph of your broker and everything it talks to (agents, MCP servers, LLMs), with edges drawn from real traffic.
- Drill into a task — one A2A turn, expanded to its iterations, tool calls, node transitions, LLM reasoning, conversation, artifacts, trace spans, and raw logs.
- Invoke the broker live — send it a message from a chat rail and watch the canvas light up node-by-node.
Three adjacent modes share the same shell: Exchange Versions, LLM Proxy, and Model Pricing (covered below). Tracer is read-only against the platform except the Invoke rail and the LLM Proxy playground, which send real requests to your endpoints, and Model Pricing, which writes model costs to API Manager when you choose to update them.

The layout
Tracer is a left sidebar plus a content column. At the top of the content column, four pill buttons switch the view mode: Broker Activity (the default), Exchange Versions, LLM Proxy, and Model Pricing.
In Broker Activity, the content splits into a network canvas on top and, once you pick a task, a task details pane below (drag the divider, or collapse either). A resizable Invoke rail sits on the right.
Picking a scope
The left sidebar is a chain of selectors — each one unlocks the next:
- 1
Business group
The scope for everything.
ALLis a valid choice but suppresses the downstream selectors. - 2
Environment
Sandbox, Production, etc. Disabled until a business group is chosen.
- 3
Broker (deployed)
The live brokers found in that environment, grouped by their parent agent network. If you see "No Brokers Activity Exists", nothing agentic is deployed there.
- 4
Activity period
The time window for the task list — 5 minutes up to 7 days, defaulting to Last 1 hour.
Time period ≠ graph window
The network graph
A hand-drawn SVG canvas (not a generic graph widget) showing four node types: broker, agent, MCP, and LLM. Edges are either design-time (declared) or runtime (drawn from actual traffic).
Getting around
- Pan by dragging the background, zoom with the scroll wheel, and drag individual nodes to rearrange.
- Keyboard: F fits the view, + / - zoom. Bottom-right buttons do the same.
- Click a node for a details card (label, type, version, framework, asset id).
- The options menu sets layout (Tree / Radial), edge style (Straight / Bent), and filters to hide Agents, MCP servers, or LLMs.
The task list
Below the selectors, the task list fills the sidebar once a broker is chosen. Each row is one task: its id, context, first tool, start/end, duration, iteration count, tools used, and status. Error tasks show a snippet inline. Click a row to open it; there's a manual refresh too.
How many tasks you see — and how much detail — depends on your entitlement (below). Without Enhanced Log Search, Tracer falls back to runtime logs: fewer tasks, capped log pages, and no trace spans.
Task details
Selecting a task opens the drill-down: a timeline/tree on the left, a tabbed panel on the right. The tabs are API status, Message, Metadata, Task story, Traces, LLM Reasoning, and Raw Log.
The left side offers Tree, List, and Graph views — but which are useful depends on the broker generation (see v1 vs v2):
- v1 brokers open on API status with a Tree/List of iterations and steps.
- v2 brokers open on Task story with a node timeline and a graph overlay of the path actually taken.
Graph execution overlay
For v2 brokers, the Graph view overlays what actually ran onto your design-time AgentScript graph: traversed edges, visited nodes, and a legend for taken / not-taken / ran-but-no-detail. If a node ran that isn't in the drawn version, you'll get a version-drift warning — the graph is the Exchange version inferred from the task's start time.
Trace hierarchy is inferred
Invoking a broker
The Invoke rail turns Tracer into a test client. It pre-fills a suggested URL from the selected broker; pick an auth mode (none / API key / basic / MuleSoft client id+secret) and load the agent card. Then chat.
- The conversation panel offers the card's example prompts and skills as one-click starters.
- Shift+Enter makes a newline; Enter sends.
- While the broker works, the canvas animates node status transitions live.
- With no URL set, it runs a local simulation instead of a real call.
Exchange Versions
Switch to Exchange Versions to compare published releases of an agent network. View a single version's files, or compare two — Tracer diffs the project files and the network topology, listing added / removed / changed nodes, and tells you plainly when there's "No differences in network topology between these versions." You can download a version's raw project zip.
LLM Proxy
The LLM Proxy mode is a test harness for Flex Gateway LLM proxies. Browse deployed proxies, then chat against one with full control over model, temperature, top-p, max tokens, and streaming. It surfaces routing decisions, the x-llm-proxy-* response headers, token usage, and a raw request/response inspector. A separate network diagram shows the request path — green for the traced route, red only when a deny list matched.
Model Pricing
Model proxies charge wallet budgets using the cost per 1M input and output tokens set on each model in API Manager's model catalog. Anypoint only lets you set those costs one model at a time, and a model with no cost counts as zero spend, so spend caps don't limit it (see Setting Model Costs for Spend Tracking). Model Pricing fills those costs in from public list prices, from OpenRouter or LiteLLM.
Pick a business group in the sidebar. This view has no environment selector, because costs don't depend on the environment (see Where model costs live). Choose a price source with the OpenRouter and LiteLLM tabs (see Choosing a price source). Each tab shows every catalog model next to that source's listing, with the current and the new price, and a status:
- Differs — the catalog price is missing or doesn't match the source.
- Up to date — the prices already match.
- Pick a match — several listings fit the model; choose one from the list.
- No match — the source has no listing for the model. Type a listing id (for example
openai/gpt-4oon OpenRouter) if you know the equivalent, or leave the model alone.
Update one row with its Update button, every differing model from one provider by choosing it in Provider and selecting Update {provider}, or the whole catalog with Update all from {source}. Updates always use the tab you're on. Bulk updates ask for confirmation first and report the result on each row.
List prices, not your contract
Choosing a price source
- OpenRouter reads OpenRouter's public model list. It has one listing per model, matched to the catalog by name. Batch and free variants are never used as a price.
- LiteLLM reads LiteLLM's community cost map on GitHub. It lists cloud models under the names the catalog stores, such as
us.anthropic.claude-opus-4-8on Bedrock, so it has the regional prices OpenRouter's single listing doesn't. Because it's maintained by the community, it can lag a provider's price change.
On the LiteLLM tab, a listing matches when its name ends with the catalog's model name and its provider is the catalog's provider. Listings from resellers never match. If several regional listings fit, the plain one wins, and if equally specific listings disagree you pick one. The listing column also shows LiteLLM's cache read, cache write and long-prompt prices. API Manager can't store those, so an update sets only input and output, but they show where wallet spend will drift from the provider's bill (see What wallets measure).
Switching tabs doesn't change anything in Anypoint. A tab shows how the catalog compares with that source, and only an update writes prices. Edit offers both sources' prices whichever tab you're on.
Where model costs live
Model costs belong to the business group, not the environment. API Manager's address for the catalog names an environment, which suggests each environment has its own prices. It doesn't:
- All environments in a group share one set of costs. Sandbox and Production show the same price and version for every model, and an update through one environment shows up in all of them straight away.
- Each business group has its own copy. The master organization and every child group hold separate costs for the same model, and they often differ.
- Nothing is inherited. Updating the master doesn't change any child group, and updating a child doesn't change the master or its sibling groups.
Price every business group
We checked this against the live API rather than assuming it. The test read every environment of a master organization and five child business groups. It then changed four models, from four different providers, in each of the six groups, and checked every environment of every group after each change. All 88 checks passed, and every price was put back afterwards.
Syncing prices to other groups
Sync to other business groups, at the top right of Model Pricing, copies the selected business group's prices to the other business groups your profile can reach. Nothing is written until you confirm:
- Review first. The dialog loads every other group's catalog and lists each one as this group → that group, with every price that would change, old and new. Groups you can't read say so and are skipped.
- Choose the groups. Untick any group to leave it alone, then confirm that you understand the sync replaces those prices.
- Only priced models are synced. A model without a price in this group never clears a price elsewhere, and models a group doesn't have are left alone. Prices someone set by hand in another group are replaced.
- No undo. Tracer keeps no copy of the old prices, so the review list is your record of them. Each group's result is shown when the sync finishes.
Setting a price by hand
Select Edit on any row to set the price yourself, without going back to Anypoint. Use it for negotiated rates, models neither source lists, or to adjust a price before or after an update. The dialog shows the current price next to each source's price, and offers:
- Save price stores the input and output prices you typed, in USD per 1M tokens. Negative numbers can't be entered. A price that looks like a per-token or per-1K figure, or one over $1,000, gets a warning before you save.
- Use OpenRouter price and Use LiteLLM price fill in that source's price for you to adjust.
- Set to zero records a deliberate price of $0, for a free or self-hosted model.
- Clear price removes the price, and the model shows Not set again.
Rows you edit are marked Set by hand, and Update all from {source} and Update {provider} skip them, so a bulk update doesn't overwrite your price. The row's own Update button still sets it to the price on the tab you're on. The marker lasts until you reload the page.
What an update changes
- Only rows marked Differs are sent. Anypoint raises a model's version on every update, even one that sets the same price, so Tracer skips models that are already up to date.
- Costs are always US dollars per 1M tokens. Anypoint stores every price in USD whatever currency is sent.
- A zero price isn't the same as no price. Zero is stored as a price, while a model with no price shows as Not set. Both count as zero spend.
- Prices can't be negative. Anypoint rejects them, and the Edit dialog won't accept one.
- Prices saved as text still read correctly. Other tools can store a price as text, such as
"10.5". Tracer reads it as a number.
Spend and wallets
Wallets, the spend and token budgets on model proxies, work the other way: each wallet belongs to one environment. The only costs an environment can see are those of its business group, so one update covers the wallets in every environment of that group. Wallets don't report spend through the API, so to check what a call cost, look at the wallet in Anypoint.
What wallets measure
Wallets measure usage in a consistent, predictable way: the tokens the provider reports, multiplied by the input and output prices set for your business group. That makes them good for setting budgets, comparing apps and stopping runaway usage.
They're an estimate based on your price list, not a copy of your provider invoice. Expect your invoice to differ because:
- Prices are the ones your business group set. Models with no price aren't counted.
- Cached input is charged at the standard input price. Providers charge much less for cached input and more for writing to the cache. Wallets use the standard input price for both, so apps that rely heavily on prompt caching use up their budget faster than their invoice suggests.
- Some pricing isn't included. Long-context rates, extras such as web search or audio, and discounts you've negotiated all leave the wallet figure unchanged.
- Only traffic through the gateway is counted. Calls made straight to the provider don't reach a wallet.
Control spending here, check charges with your provider
Permissions
The entitlement wall
The biggest surprise for new users. Full-fidelity tracing depends on Enhanced Log Search, which requires the Anypoint Advanced package or a Titanium subscription. Tracer probes this per organization, so switching business groups can change what you see.
Without the entitlement, Tracer still works
For full task visibility, the deployment also needs the INSECURE-LOGGING monitoring category enabled in Runtime Manager, and the broker's Object Store keys must not have expired.
Gotchas
- No exports. There's no canvas image download and no task CSV/JSON export. To keep a trace, copy from the Raw Log tab. (The only screenshot feature anywhere is the bug-report widget.)
- Object Store data expires. Task story and LLM reasoning vanish once keys TTL out (24h default). An empty panel on an old task is retention, not a failure.
- v1 and v2 look different. Same UI, different tabs and default view. Don't assume a feature is missing.
- LLM reasoning extraction is heuristic. For older brokers it's scraped from serialized Object Store payloads — partial or noisy output is expected.
- Only the first instance is queried. Multi-instance brokers may under-report tasks.
- A2A version mismatch in Invoke produces confusing broker errors — match 0.3 vs 1.0 to the broker.
- Nothing is URL-addressable but the mode. Help can deep-link to Broker Activity / Exchange Versions / LLM Proxy / Model Pricing, but not to a specific broker or task.
