Saltar al contenido
ES EN

AI for developers: the week Nov 9–15

Spend caps, read-only agent workflows, TODO cleanup and GitOps audits all point the same way this week: guardrails for agents are now part of the job.

Read this week's developer coverage together and one pattern jumps out: the interesting news is no longer what agents can do, but how to stop them doing too much. GitHub shipped its agentic workflows to general availability, Google Cloud added spend caps for AI services, and Google's own research put numbers on how forgotten TODO comments pull code agents off task. Each piece is a different angle on the same problem.

Start with the plumbing. GitHub's agentic workflows begin with read-only permissions and push write operations into separate jobs with their own tokens. That is a sensible default. The catch, according to the project FAQ, is that the protection does not automatically cover direct write permissions or custom jobs the workflow author adds. In other words, the isolation is real, but you can punch a hole through it with one line of YAML, and we suspect many teams will.

Illustration of an AI coding agent inside a fenced box with a budget meter and a read-only lock
The common thread this week: fence the agent, cap the spend, and keep the context clean.

Money is the second fence. Google Cloud's spend caps in Cloud Billing Budgets block new requests once a service's estimated monthly spend reaches the budget, without deleting anything. It covers the Gemini API, Gemini Enterprise Agent Platform, Cloud Run and Cloud Run functions. Because the comparison uses estimated spend, you want a margin below your real limit rather than a cap set exactly at it. That is a defence against a runaway agent loop, not a substitute for watching your bill.

The third fence is the least obvious: your own repository. The finding in the TODO cleanup guide is that distracted sessions cost 70% more steps, 77% more tokens and 105% more waiting time. Google also notes its cleaning agents sometimes had to read years of history to confirm that a TODO and its intent were both dead. So this is review work with tooling help, not a script you run and forget.

There is a quieter lesson in the Argo CD audit. LinkedIn runs over 250 clusters and 100,000 deployments a week, and cut deployment infrastructure failures from 925 to 63 by replacing its GitOps engine. Most of us are nowhere near that scale, but the two checks the guide proposes are cheap and read-only, which is exactly the right posture before changing anything. We would run them before arguing about migration.

Our reading: the week's practical takeaway is a short checklist. Keep agent tokens read-only unless a step truly needs to write. Put a hard cap on every metered AI service. Prune stale TODOs before they become accidental prompts. And treat any claim of safety, including the ones in vendor FAQs, as a starting configuration to verify. The tooling is getting better at containment, but the containment still has to be configured by someone who has read the fine print.

Related articles from LaRebelión

Guide to Cleaning Up Obsolete TODOs and Preventing Them from Distracting Your AgentsGuide to Cleaning Up Obsolete TODOs and Preventing Them from Distracting Your AgentsHow to Configure GitHub Agentic Workflows with Budget Controls and IsolationHow to Configure GitHub Agentic Workflows with Budget Controls and IsolationGuide: How to Set a Spend Cap in Google Cloud for Your AI ProjectsGuide: How to Set a Spend Cap in Google Cloud for Your AI Projects

Original source: La Rebelión

Article generated with AI.larebelion

Byline

· Chief editor · English edition · London

“A forgotten TODO is now an instruction your agent might follow, and nobody on the team wrote it for that.”

This week’s news

Comentarios