AWS Builder Center

Full Visibility into Your Coding Agents: Amazon CloudWatch Coding Agent Insights

Amazon CloudWatch Coding Agent Insights centralizes OpenTelemetry telemetry from Claude Code, OpenAI Codex, and GitHub Copilot into pre-built dashboards, reporting token consumption, cost, developer productivity, and latency segmented by department, cost center, team, and user.

Senior Specialist SA, GenAI/ML
Your developers are already using coding agents. The harder question is what that usage costs, which teams are getting value from it, and whether the next license expansion is justified.
 Amazon CloudWatch Coding Agent Insights  answers those questions with ready-to-use dashboards that centralize telemetry from  Claude Code ,  OpenAI Codex , and GitHub Copilot. This post covers how the telemetry gets to CloudWatch, what the dashboards show, and how organizational attributes turn aggregate numbers into per-department chargeback.

Why This Matters

Coding agent adoption tends to outpace the ability to measure it. A platform team enables Claude Code for one squad, the squad tells another, and within a quarter the agent is running across three departments with no shared view of consumption. Token by token, session by session, spend grows.
The consequences show up in specific conversations. An engineering leader is asked to justify next year's budget and has no per-team consumption data. A FinOps analyst needs to attribute agent spend to cost centers and finds a single undifferentiated line item. A platform team is choosing between two models and has no measured cost-to-output comparison to reason from, so the decision gets made on preference.
Each of these is a reporting problem, not a technology problem. The telemetry already exists.

How Coding Agent Insights Works

Coding agents natively emit  OpenTelemetry  (OTel) metrics: token consumption, per-turn latency, tool calls, API requests, and approval decisions. Coding Agent Insights receives that telemetry through the CloudWatch native OTLP metrics endpoint and renders it in pre-built dashboards under GenAI Observability → Coding Agent Insights in the CloudWatch console.
There is nothing to create or import. The dashboards appear automatically once agents send data in the expected shape. Your only task is configuring each agent to publish its OTel metrics to CloudWatch.

Prerequisites

You need an AWS account with CloudWatch enabled, an agent version that supports OTel export, and credentials permitting OTLP metric ingestion. Before you roll out to more than a handful of developers, settle on the taxonomy for your organizational attributes — the values you choose for  department ,  cost_center , and  team  are what every subsequent filter and chargeback report depends on, and reconciling inconsistent values after the fact is tedious.

Supported Agents

Agent
Setup documentation
Claude Code
 Set up OpenTelemetry for Claude Code 
—
 Bearer Token 
or
 Enterprise 
OpenAI Codex
 Set up OpenTelemetry for OpenAI Codex 
—
 Bearer Token 
or
 Enterprise 
GitHub Copilot
 Set up OpenTelemetry for GitHub Copilot 
— Bearer Token

Choosing a Setup Mode

Bearer Token mode is the fastest path for an individual developer or a small team. The agent sends OTLP metrics directly to the CloudWatch endpoint with an  Authorization: Bearer  header, no collector is required, and identity and organizational attributes travel in the  OTEL_RESOURCE_ATTRIBUTES  environment variable.
Enterprise mode differs by agent. For Codex it is an enterprise rollout that federates identity through your corporate SSO provider — Okta, Microsoft Entra ID, Auth0, or AWS IAM Identity Center — and centralizes Amazon Bedrock access. For Claude Code it is implemented with the self-hosted Claude apps gateway, where developers sign in with your corporate identity provider and the gateway governs model access and relays telemetry. The gateway supports Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry as upstreams.
The practical difference between the two is attribution quality. In Bearer Token mode each developer sets their own resource attributes, so a typo in  department  files that person's consumption under the wrong bucket. Enterprise mode sets those attributes centrally, which is what you want in place before anyone uses the numbers for chargeback.

Organizational Dimensions

Every dashboard can be segmented and filtered by the organizational attributes your agents send as OTel resource attributes:
  • Organization — the company or organization
  • Department — engineering, product, QA
  • Cost Center — for chargeback
  • Team — a specific team
  • User — individual developer, by email
  • Role — developer role
  • Model — the AI model in use
  • Location — geographic location

Exploring the Claude Code Dashboard

The screenshots in this section come from a demonstration environment with synthetic data. The following screenshot shows the Claude Code tab.
Claude Code general dashboard in Coding Agent Insights

The executive summary answers the first question most leaders ask: what are we spending, and who is using it. It reports total tokens and total cost in USD for the selected period alongside active users, session count, and active hours, which is enough to derive per-developer and per-session cost without exporting anything. The cache hit rate reads 99.4% here. Because cached reads bill at a fraction of input tokens, a declining hit rate translates fairly directly into higher spend, which makes it a reasonable alarm target.
The usage and cost panels break that total down along two axes. A time series shows whether consumption is rising because adoption is growing or because prompts are becoming less efficient — two situations that call for opposite responses. Department-level views of both tokens and cost give you the figures a chargeback conversation needs. The token type breakdown separates input, output, cache read, and cache creation, which matters because those four categories are priced differently.
Cost only means something set against output. The developer productivity panels report lines of code generated, plus commits, pull requests, and active hours per department. These are adoption and throughput indicators rather than performance measures. Lines of code has never been a good proxy for value, but a team producing commits and pull requests at a steady rate has clearly integrated the agent into its workflow, and a team with high token consumption and flat commit activity is worth a conversation.
The code editing panels show where the agent is being used and how much of its output survives review. Generation is split by language across Python, TypeScript, JavaScript, SQL, and Bash, and edit decisions record accept and reject counts. The accept rate is the closest thing on this dashboard to a quality signal: a team rejecting most proposed edits is one where the agent is not yet well matched to the codebase.

Filtering by Organization

The panels give you aggregates. Filtering is what turns them into answers about a specific part of the organization. The following screenshot shows the view narrowed to a single developer.
Filters applied in Claude Code — organization, department, cost center, user, and rol

The view is scoped to organization  octank-corp , department  engineering , cost center  cc-1234 , user  ffagas@octankgoods.com , and role  software-engineer . The dashboard recalculates immediately, which supports the three conversations from earlier: an engineering leader reviewing departmental spend, a FinOps analyst producing chargeback figures by cost center, and a manager comparing throughput across teams.

Exploring the OpenAI Codex Dashboard

Codex reports on the same cost and attribution axes, and adds two dimensions Claude Code does not: request outcomes and latency percentiles. The following screenshot shows the Codex tab.
Codex general dashboard in Coding Agent Insights

The overview covers total tokens, at 132K in this environment, along with API requests, active users, conversation turns, and a 74.2% cache hit rate. Conversation turns is the useful denominator here, because Codex work is measured in exchanges rather than sessions.
Token usage follows the same time series and department breakdowns, with one addition to the type split: reasoning output is reported separately from output. On reasoning models it is often the larger of the two, and separating them tells you whether a cost increase came from longer answers or from more deliberation.
The API and tool activity panels are where operational problems surface. Requests are grouped by outcome, so a rise in failures becomes visible before developers report it. Tool calls are broken out by tool —  apply_patch ,  exec_command ,  tool_search , and  update_plan  — which describes how the agent is working rather than just how much it is being used. Approval decisions track how often developers accept the actions Codex proposes.
Latency is reported at p50, p90, and p99 for conversation turns, time to first token, and tool calls, with p90 API request latency also split by department. Percentiles are the right view for developer tooling. A p99 turn latency of several seconds is the experience that sends people back to writing code by hand, and it disappears in an average.

Filtering by Model and Environment

Codex view offers the same organizational filters as Claude Code plus Environment, which separates development, staging, and production, and Model, which isolates consumption by a specific model. The following screenshot shows the dashboard filtered to one model.
Filters applied in Codex — model openai.gpt-5.6-luna selected

With the view scoped to  openai.gpt-5.6-luna , you can compare cost, latency, and accept rates against another model such as  openai.gpt-5.4  on the same workload, and route each workload to whichever model gives the better cost-to-output ratio. This is the measurement that model selection debates usually lack.

Conclusion

Coding Agent Insights closes a specific gap: it gives you operational visibility into coding agent adoption at organizational scale without building or maintaining dashboards. The dashboards appear on their own once telemetry arrives, they cover Claude Code, Codex, and Copilot in one console, and the organizational attributes make real chargeback and evidence-based expansion decisions possible. Because the underlying metrics are queryable with PromQL, you can also build custom views and proactive alarms on top of them.

Getting Started

  1. Pick your agent and follow the  setup guide .
  1. Configure organizational resource attributes such as  department ,  cost_center , and  team  so segmentation works from the first metric.
  1. Open the CloudWatch console and navigate to GenAI Observability → Coding Agent Insights.
  1. Set per-department token budget alerts with PromQL.
Coding Agent Insights is available in all AWS commercial Regions except Middle East (UAE), Middle East (Bahrain), and Israel (Tel Aviv). Standard CloudWatch OpenTelemetry metric ingestion pricing applies; see  CloudWatch pricing  for details.


Francisco Fagas is a Senior Specialist Solutions Architect for Generative AI at AWS, based in San Juan, Puerto Rico, covering LATAM NOLA enterprise accounts. With more than 20 years of experience in IT, he combines technical expertise across cloud architecture, application development, and AI/ML to help enterprise customers adopt generative AI and accelerate innovation through agentic AI, large language models, and modern development practices.


Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article