Prewarm the environments agents work in, and trim what they log
Two agent-operations stories today: Shopify's warm pool of build machines for mobile agents, and a worked example of cutting Claude Code telemetry by about half. The engineering picks are Python's expedited security releases for the eMudhra CA incident and AWS's root volume replacement for patching stateful EC2 hosts.
AI & Agent Engineering
Agent infrastructurevendor-reported
Shopify gives its Slack agent remote Macs, and its share of merged mobile pull requests rises from 6% to 61%
Shopify says three or four parallel agent workflows were enough to max out a laptop, so it built Remote Dev: a fleet of remote Macs plus a Linux VM for Android emulators, with each River session claiming its own environment. A pool of warm Macs with main checked out and caches built replaced a setup that took over 30 minutes, cutting it to a few minutes.
Why it matters
The bottleneck for agent throughput was the environment, not the model. The post also says River initially created 6% of merged mobile pull requests and reached 61%, and credits a skill, a CLI wrapper and tmux panes for making the agent use the remote machine. Figures are Shopify's own.
Do this week
Time how long an agent takes to get from task request to a running build in your repo. If it is more than a few minutes, try a small pool of prewarmed environments that are replaced after each session.
CNCF tutorial: Collector rules roughly halve Claude Code log and trace bytes and cut 300 metric series to 12
The author, from Bindplane, ran Claude Code on a small task with every content flag on and got 181 telemetry records and 289 KB of JSON. Collector filters, attribute deletes and truncation roughly halve log bytes and trace bytes, and deleting session and user attributes takes 300 metric series for 25 sessions down to 12.
Why it matters
Defaults suit debugging one session, not a fleet: tool output is stored in both traces and logs, and per-session metric attributes create new series for every session. Numbers come from one small run by a vendor employee.
Do this week
List the questions your team asks of agent telemetry, such as cost per model or failing tools, and drop every field no dashboard or alert reads. Hash or delete user and session identifiers on metrics first.
Python 3.15.1, 3.14.9 and 3.13.17 ship pip 26.2.2 for the eMudhra CA incident; virtualenvs need upgrading
The Python release team issued expedited security releases of 3.15 and 3.14 and a special security release of 3.13. All three bundle pip 26.2.2, which includes certifi 2026.10.19, to address the eMudhra and emSign CA security incident. The post tells users to upgrade pip and uv in all existing virtual environments.
Why it matters
This follows the Node.js fix for the same incident. Upgrading the interpreter alone does not help virtualenvs that carry their own pip and certifi, and the team notes it does not normally update pip in a release like 3.13 because of backward compatibility.
Do this week
Inventory Python base images and virtualenvs in CI and production, rebuild them on the new releases, and check which certifi version each one resolves, including copies bundled by other tools.
AWS shows how to patch or upgrade a stateful EC2 host by swapping a prepared root volume in one API call
AWS describes preparing and validating a replacement root volume on a separate instance, then calling create-replace-root-volume-task. The instance reboots and keeps its ID, private IP, network interfaces, Elastic IP and EBS data volumes. Constraints include same Availability Zone, matching CPU architecture and boot mode, and ENA and NVMe drivers on the new volume.
Why it matters
For hosts that hold identity and attached resources, this limits downtime to a reboot and keeps a rollback path if you keep the original volume. The post warns that a fresh AMI snapshot can re-run user data and change SSH host keys.
Do this week
Pick one stateful host that you currently replace by hand, and rehearse the swap on a non-production copy, including the rollback to the original root volume, before the next OS patch window.
Measure how long an agent waits for a working environment
Pick one repository and have an agent start from a fresh clone on a clean machine. Record the minutes until the first successful build and test run, then repeat with a prebuilt image or cache.
Why. Shopify's reported gain came from warm environments, not a better model. This shows whether setup time is limiting how many agent tasks you can run in parallel.
Measure. Minutes from task request to first green test run, cold versus warm, and the number of parallel sessions one machine handles before builds slow down.
Classic / essential read
Yagni
Martin Fowler · 10 min
Fowler explains that building a presumptive feature now has a cost of build, cost of delay, cost of carry and cost of repair, and that the saving only holds if the code stays easy to change.
One lesson. Before building for a requirement that is months away, count what the extra code costs to carry, and invest in making the code easy to change instead.
Throughput and cost often hide in the setup around the work: the environment an agent waits for, the telemetry it emits, the certificates under your runtimes. Pick one this week and put a number on it.
Get this every morning.
Free. Same issue for everyone. One click to unsubscribe.