Skip to content
octic.ai
Sign up

Trace an AI agent run, end to end.

Coming soonThe agent run explorer will trace each run of the AI agents you build: every model call and tool call, in order, with its latency, tokens, and cost, beside the AI agent’s owner and trust state. Here’s a preview.

Sign upBook a demo

One run of support-triage-agent, step by step

run_4d1e: 5 steps, 6.8 s, $0.11.

  1. Step 1: Model call, Anthropic API

    1.64 s

    3,420 tokens$0.0198

  2. Step 2: Tool call, mcp-jirasearch_issues

    0.64 s

    No tokens

  3. Step 3: Model call, Anthropic API

    1.76 s

    6,360 tokens$0.0371

  4. Step 4: Tool call, mcp-jiraadd_comment

    0.48 s

    No tokens

  5. Step 5: Model call

    Anthropic API

    2.28 s

    Input
    The ticket, ENG-2291, and the new comment
    Output
    Reply drafted, ticket routed to @platform
    Started
    +4.52 s
    Latency
    2.28 s
    Tokens
    8,120 in, 520 out
    Cost
    $0.0531
    Through
    Anthropic SDK 1.3.0
    Reached
    Anthropic API
    Tokens
    8,120 in, 520 out
    Cost
    $0.0531
    Through
    Anthropic SDK 1.3.0

In this preview on Acme Corp's map, ip-172-31-67-241 runs support-triage-agent, an approved AI agent owned by @support-eng. Its run run_4d1e would show 3 model calls to the Anthropic API and 2 tool calls to mcp-jira, the slowest starting at +4.52 s.

The Octic SDK will report every run

When the Octic SDK ships, you’ll add it to the application that runs your AI agent. Each run will then be reported as it happens: every model call and tool call, in order, with a summary of what went in and what came out.

Coding agents won’t need it. Claude Code and Codex sessions are reported by the sensor today, with no SDK, and you can see them now in Agentic Coding Insights.

Preview: an AI agent instrumented with the Octic SDK will report each run; a coding agent's sessions come from the sensor today.
  • Preview: support-triage-agent, an AI agent in Acme Corp's own application, runs on ip-172-31-67-241. Instrumented with the Octic SDK (coming soon), it will report each run.
  • Its run run_4d1e would be reported with every step: 3 model calls and 2 tool calls.
  • Claude Code on eng-ws-14 needs no SDK: the Octic sensor reports its sessions today, and they are shown in Agentic Coding Insights.

Every step, timed and costed

Each step will carry when it started and how long it took. Model calls will also carry their tokens in and out, and what those cost. The steps will add up to the run, so a slow or expensive run will point to the step that made it so.

Runs will roll up to their AI agent, so you’ll see which AI agents cost the most.

run_4d1e by step: time, tokens, and cost

Model calls took 5.68 s of the run’s 6.8 s and all $0.11. The 2 tool calls took 1.12 s and no tokens.

Time6.80 s

Time by step: step 1, 1.64 s; step 2, 0.64 s; step 3, 1.76 s; step 4, 0.48 s; step 5, 2.28 s.

Tokens18,420

Tokens by step: step 1, 3,420; step 3, 6,360; step 5, 8,640.

Cost$0.11

Cost by step: step 1, $0.0198; step 3, $0.0371; step 5, $0.0531.

The 3 model calls reached the Anthropic API: 5.68 seconds, 18,420 tokens, $0.11. The 2 tool calls reached mcp-jira (search_issues and add_comment): 1.12 seconds, no tokens.

Each run on the same map

A run will link back to its AI agent on the map, so the agent’s owner and trust state will sit beside the trace, along with the owner and trust state of everything the run reached.

The host the sensor watches is already on the map, so what the SDK reports will line up with what the sensor finds, down to a package nobody has reviewed that the run never used.

Preview: the path run_4d1e took through the map: the agent, its owner and trust state, and what it reached.
  • ip-172-31-67-241, a Linux host with the Octic sensor, runs support-triage-agent and has two packages: Anthropic SDK 1.3.0 (approved) and openai 1.51.0 (unreviewed).
  • support-triage-agent is owned by @support-eng and approved. Its run run_4d1e went through Anthropic SDK 1.3.0.
  • The run reached the Anthropic API (owner @eng-leads, approved) with 3 model calls, and mcp-jira (owner @platform, approved) with 2 tool calls.
  • It did not use openai 1.51.0.
Show on the map

run_4d1e reached 2 services, both approved. openai 1.51.0, unreviewed and on the same host, went unused.

Where this sits in Observe

Agentic Coding Insights measures the coding agents your engineers use, from the sensor, today. The agent run explorer, coming soon, will trace the AI agents you build, from the SDK. Both land on the same map.

Back to Observe

Start with what ships today

Agent run tracing is coming soon. Install a sensor on a Linux or Windows host today and Claude Code and Codex sessions start landing on the map, along with the AI running on it.

Sign upBook a demo

Or join the Design Partner Network.