Agents' Output Now Shows Up on the Management Dashboard — CloudWatch Starts Measuring Claude Code, Codex, and Copilot Side by Side
Announced by AWS on July 20, CloudWatch Coding Agent Insights lines up the consumption, cost, and output of Claude Code, Codex, and Copilot on a single screen. This piece weighs both sides — the benefit of measuring adoption gains in hard numbers, and the danger of repurposing individuals' line coun
Coding agents are no longer a "set it and forget it" tool. The more they get adopted, the more the next question becomes an operational one of visibility: who is using how much, and what are they producing? On July 20, 2026, AWS answered that question head-on with a new feature, Amazon CloudWatch Coding Agent Insights — a dashboard that puts usage data for Claude Code, OpenAI Codex, and GitHub Copilot on the same screen as your existing CloudWatch operational data.
Putting numbers to adoption gains that used to be described by gut feel
Until now, the impact of coding agents tended to be described in terms of gut feel — "it feels like development got faster." Coding Agent Insights collects that impact as metrics sent via OpenTelemetry (an open instrumentation standard) and presents it along three axes: consumption, cost, and output. The idea is to let engineering leaders decide — with data rather than guesswork — which teams to expand access for, where development is accelerating, and how to right-size each department's token budget.
The dashboard is split into four tabs
The screen is dedicated to coding agents, and you can switch between views depending on what you want to see.
| Tab | What it shows |
|---|---|
| Executive Summary | Total tokens, total cost, active users, session count, cache hit rate, uptime |
| Usage & Cost | Token consumption over time, cost by department and team, and a breakdown by token type — input, output, and cache reads/writes |
| Developer Productivity | Lines added and removed, number of commits, actual working time, number of pull requests |
| Code Editing | Language distribution, the accept/reject ratio for suggested edits, and which editing tools were used |
It is available in all AWS commercial regions except Israel (Tel Aviv) and the Middle East (UAE and Bahrain), and pricing follows CloudWatch's standard metric ingestion rates.
Where does the data come from? The key is the "Claude apps gateway"
The numbers originate from the OpenTelemetry metrics each agent emits. For Claude Code, you can collect telemetry with no additional instrumentation code by routing it through the Claude apps gateway (an in-house control plane) that AWS announced on July 15. This gateway is a stateless container bundled with the Claude Code CLI binary, and it can run on ECS, EKS, or EC2. Beyond simply collecting numbers, it also handles governance such as the following:
- Internal single sign-on via OpenID Connect (with settings applied per identity-provider group)
- Distributing which models and tool permissions are available as defaults that can't be overridden locally
- Capping spend with daily, weekly, and monthly limits set per organization, group, and individual
- Routing to multiple execution backends such as Bedrock, Claude Platform on AWS, Google Cloud, Microsoft Foundry, and the Anthropic API
Configuring the Claude Code side is done through environment variables
If you just want to send metrics from your local Claude Code, you can do it entirely with environment variables, without the gateway. You enable telemetry and point the destination at a CloudWatch endpoint.
- Enable telemetry with
CLAUDE_CODE_ENABLE_TELEMETRY=1 - Specify the destination with
OTEL_METRICS_EXPORTER=otlpandOTEL_EXPORTER_OTLP_ENDPOINT(https://monitoring.<region>.amazonaws.com) - Attach
user.id/team.id/cost_centerand the like toOTEL_RESOURCE_ATTRIBUTESto associate consumption with a specific person or team
After configuring it, run a short session, and if metrics such as claude_code.token.usage arrive in CloudWatch, connectivity is confirmed.
From "just using it" to "deciding how to allocate it"
The business significance goes beyond cost visibility. Right-sizing token budgets per department, shifting capacity toward the teams that are delivering results, and preventing runaway usage with spending caps — you can put agent usage on the footing of allocation decisions rather than hunches. At a time when generative AI usage tends to balloon as an "invisible expense," being able to line up consumption and output on the same screen has real practical value.
The danger of numbers turning into surveillance
On the other hand, some warn that caution is needed around the fact that the Developer Productivity tab can show "lines added," "commit count," and "pull request count" at the individual level. If you bring lines of code or commit counts into individual evaluations as productivity metrics, the metric itself becomes the goal, which can invite padding and unnecessary changes (the age-old problem that when a measure becomes a target, it ceases to be a good measure). These aggregates are meant to capture trends across teams and departments; using them as a yardstick to grade individuals risks a hidden cost behind the convenience — the surveillance and chilling of developers. Deciding in advance what to measure, who gets to see it, and what it will not be used for looks like what will separate a good rollout from a bad one.
References: Amazon CloudWatch announces coding agent insights (AWS) / Analyzing Claude Code usage with CloudWatch and OpenTelemetry (AWS Cloud Operations Blog) / AWS Ships Claude Apps Gateway as Self-Hosted Control Plane (InfoQ) / Set up OpenTelemetry for Claude Code (Amazon CloudWatch Docs)
