Tech Radar — Agents move into production, and the discipline follows
This edition: Claude Sonnet 5 makes agents cheaper to run, AWS gives every agent its own micro-VM, Copilot lands security autofixes in Azure DevOps, Meta eyes the cloud market, and platform teams grow up.
The thread running through this week is the same one: AI agents are no longer demos, they are workloads. And the moment something becomes a workload, the boring questions arrive — what does it cost, where does it run, who isolates it, and who reviews what it ships. Here is what I would put in front of an engineering leadership team this week, and why each one matters beyond the headline.
Anthropic ships Claude Sonnet 5 as a cheaper way to run agents
Anthropic launched Claude Sonnet 5 with stronger agentic capabilities and lower pricing, positioning it as a cheaper alternative to Opus, GPT-5.5 and Gemini Pro for teams running agents at volume (TechCrunch).
Impact for companies: the price of the frontier is no longer the price you pay for most work. A cheaper mid-tier model that is "good enough" for high-volume agentic tasks changes the unit economics of anything you were hesitating to automate because the token bill did not close.
Risks and opportunities: the risk is quietly standardizing on one vendor's pricing curve and waking up locked in. The opportunity is a model portfolio — a frontier model for hard reasoning, a cheaper tier for the 80% that is routine — chosen per task, not per hype cycle.
My take: run the math on cost per completed task, not cost per million tokens. A model that is 3x cheaper but needs two retries is not cheaper. Benchmark on your own workloads before you migrate anything that matters.
AWS gives every agent its own micro-VM
AWS launched Lambda MicroVMs, a serverless primitive that runs each user session or AI agent inside its own Firecracker virtual machine — hardware-level isolation, snapshot-based fast launch, and state held for up to eight hours. Community analysis pegs the minimum at around $3.03/day, roughly 9x Fargate spot pricing (InfoQ).
Impact for companies: this is the infrastructure answer to the runaway-agent problem. If an agent executes untrusted code or acts on user input, per-session hardware isolation stops one tenant's blast radius from becoming everyone's incident.
Risks and opportunities: the risk is the price tag — 9x is not a rounding error at fleet scale, and it is easy to reach for isolation you do not need. The opportunity is a cleaner security posture for the genuinely risky workloads, without hand-rolling your own sandboxing.
My take: isolation is a spectrum, not a switch. Reserve micro-VMs for code you do not trust; keep cheaper compute for code you do. The engineering decision here is classifying your workloads honestly, not defaulting everything to the most expensive tier.
Copilot Autofix lands security remediation inside Azure DevOps
Microsoft opened a public preview of Copilot Autofix for GitHub Advanced Security on Azure DevOps, bringing AI-generated vulnerability fixes to teams working in Azure Repos (InfoQ).
Impact for companies: the bottleneck in application security was never finding vulnerabilities — scanners have flooded us with findings for years. It was fixing them. Moving suggested remediations into the pull request is where security debt actually gets paid down.
Risks and opportunities: the risk is a false sense of safety — an AI-proposed fix that closes the scanner alert without closing the actual hole. The opportunity is turning a stale backlog of "known issues" into reviewed, mergeable diffs.
My take: autofix is a drafting tool, not an approver. Every suggested patch still goes through human review and the test suite. Used that way, it shortens the distance between "we know about this CVE" and "it is shipped" — which is exactly where most orgs bleed time.
Meta wants to sell you cloud compute
Meta is reportedly building a cloud infrastructure business, selling access to its AI compute and models — putting it in direct competition with AWS, Google Cloud and Azure, and echoing the way SpaceX monetizes excess capacity (TechCrunch).
Impact for companies: more serious cloud suppliers is good for buyers. Competition on GPU availability and price is the most direct lever on the AI infrastructure bill that has quietly become a top-three line item at a lot of companies.
Risks and opportunities: the risk is chasing a cheap new entrant and inheriting an immature platform, thin tooling and unproven reliability. The opportunity is real negotiating leverage and a credible multi-cloud story for the compute-heavy workloads.
My take: a new vendor is a procurement lever before it is an architecture decision. Keep the AI-heavy parts of your stack portable enough that price competition works in your favor — but do not migrate production onto a day-one platform to save a few points.
Platform teams grow up: from projects to products
InfoQ documented a team that shifted from project-thinking to product-thinking once its internal platform outgrew a single team — moving to self-service, API-driven, multi-tenant infrastructure with clear ownership, after hitting the classic limits of one-off deliveries and weak feedback loops (InfoQ).
Impact for companies: most internal platforms die of the same disease — treated as a backlog of tickets instead of a product with users, a roadmap and adoption metrics. The reframe is organizational, not technical.
Risks and opportunities: the risk is a "platform" nobody chooses to use, which quietly becomes a tax. The opportunity is genuine leverage: every hour invested compounds across every team that self-serves instead of filing a request.
My take: if your platform team cannot name its users and its adoption rate, it is running projects, not a product. Give it a product owner, measure self-service adoption, and treat internal developers as customers who can say no.
Securing AI-accelerated development
A QCon talk mapped the concrete vulnerabilities inside the agent ReAct loop — context, reasoning and tool execution — and the defense-in-depth patterns for production: memory-poisoning mitigations, LLM-as-a-judge critics, and MAESTRO threat modeling (InfoQ).
Impact for companies: as agents gain the ability to call tools and act, the attack surface moves from "bad output" to "bad actions." That is a governance question that lands on engineering leadership, not just security.
Risks and opportunities: the risk is shipping autonomous capability faster than the guardrails around it. The opportunity is that the patterns are now known — this is no longer uncharted territory, it is discipline you can adopt.
My take: treat an agent with tool access like a service with production credentials — least privilege, audit logs, and a human gate on anything irreversible. Productivity from AI is real, but it is only trustworthy when the controls ship alongside it.
The trend to watch
Every item this week is really about the same maturation: the industry is done being amazed that agents work and has started asking what it costs to run them safely. Cheaper models, per-agent isolation, security fixes in the PR, more compute suppliers, product-grade platforms — these are the unglamorous foundations that turn a demo into a dependable system. The teams that win the next year will not be the ones with the flashiest agent. They will be the ones who wrapped it in cost controls, isolation and review before their competitors had to learn why that matters.