Agent Runs

See everything your agent did. Prove every step was safe.

One run groups an agent's tool calls and model turns into a single trace: steps, decisions, cost, latency, and models in one place. Other platforms show you what your agent did. AxioRank shows you that plus whether each step was safe, and hands you signed proof.

Open the console Read the docs

Runs, sessions, and kill-chain detection on every plan. Per-run cost and latency on Pro and Team.

run 7f3a91c24 steps · $0.021 · 1.9 s

0model.completiongpt-4o · 1.2 s · $0.021allow

1vault.read82 msallow

2web.fetch610 ms · taintedallow

3http.postexfil attemptdeny

Step 3 was blocked, and the whole run is verifiable offline.

20+

framework integrations, one line each

1 trace

tool calls and LLM turns, stitched

Signed

every step offline-verifiable

What you see

The run is the unit, not the log line.

Each run rolls its steps up into one row: the decision mix, total cost and tokens, the models used, duration, and p95 step latency. Open a run and walk the timeline step by step, from the user's intent to the final call.

Step timeline

Every tool call and model turn in order, each with its decision, risk, signals, latency, cost, and redacted payload.

Sessions

A session groups the runs of one conversation. Jump from a run to its siblings with one click.

Your metadata

Tag calls with environment, feature, or your own request id. Tags surface on the timeline, redacted at write.

Cost per run

What did this run cost, in dollars and tokens, across every model it touched? One number, per run.

Kill-chain detection

Multi-step attacks surface on the run itself: the contributing steps light up, in sequence, with the alert attached.

Cross-agent lineage

When tainted data flows between agents, the runs link to each other. Follow the data, not just the log.

Drop it in

One trace id. Tool calls and model calls, stitched.

Start a trace in the SDK and every guarded tool call joins the run automatically. Send the same ids as headers on AI Gateway model calls and the LLM turns land in the same timeline, next to the tools they drove.

import OpenAI from "openai";
import { AxioRank } from "@axiorank/sdk";

const axio = new AxioRank({ apiKey: process.env.AXIORANK_KEY });
const t = axio.trace({
  intent: "Refund order #4821",
  sessionId: conversationId,          // groups this run with its siblings
  metadata: { env: "prod", feature: "refunds" },
});

// Model calls through the AI Gateway join the SAME run:
const openai = new OpenAI({
  baseURL: "https://www.axiorank.com/api/proxy/v1",
  defaultHeaders: { "X-AxioRank-Key": process.env.AXIORANK_KEY, ...t.headers() },
});

await t.enforce({ tool: "stripe.refund", arguments: { order } }); // step 0, governed

The difference

Agent observability, from the thing that governs the agent.

A gateway that only observes can tell you a run was slow and what it cost. It cannot block step 3, hold it for approval, or prove any of it happened. AxioRank can, because the run view reads the same governed rows the policy engine wrote.