State of AI Engineering

What The State of AI Engineering Report Reveals About AI in Production

AI engineering has moved past experimentation, but are your observability practices keeping up? Datadog analyzed LLM telemetry from 1,000+ customers to reveal what's actually happening at scale: model adoption shifts, hidden token costs, and what the rise of agentic frameworks means for reliability. Download the report to benchmark your AI stack against production reality.

Atalssian move to OpenTelementry - Atlassian is migrating its metrics platform from StatsD to OpenTelemetry while keeping the existing interface, allowing thousands of services to move without an immediate rewrite. The new pipeline processes metrics from around 100,000 hosts across 14 regions, using separate OpenTelemetry Collector stages for different tasks. The migration also reduces CPU usage and lays the groundwork for moving application instrumentation to OpenTelemetry SDKs. Read more.

OpenAI’s answer to Jev! - OpenAI has introduced a Decision API powered by its Luna model, designed to return structured decisions with confidence scores instead of normal chatbot responses. The API is aimed at tasks like classification, routing, and choosing an agent’s next action, with responses designed to be much faster than regular LLM calls. It is also presented as an answer to Jev, giving developers a simpler way to build systems that need reliable decisions rather than generated text. Read more.

AI Is Moving From Chatbots to Agents: OpenAI and Google Take Different Paths

AI companies are no longer competing only to build models that answer questions. They are also working on systems that can complete tasks, operate software, and support entire workflows. OpenAI’s Dots, GPT-6.1 Sol, and Google DeepMind’s Gemini 4 Argon show three parts of this shift: AI agents that work continuously, models that make advanced capabilities more affordable, and models designed for complex real-world work. Although they overlap, each has a different purpose.

OpenAI Dots: AI That Keeps Working

OpenAI’s Dots are always-on AI agents designed to work toward goals even when users are not actively chatting with them. Powered by GPT-6 Astra, each dot has its own cloud computer and can connect to more than 4,000 apps. It can investigate bugs, prepare code fixes, update documents, and carry work forward between conversations. Users can review its progress, set permissions, and require approval for sensitive actions. Dots are therefore more than a model: they combine AI intelligence with tools, computing resources, and ongoing task management.

GPT-6.1 Sol: More Capability at Lower Cost

GPT-6.1 Sol focuses on making advanced AI more affordable to run. OpenAI says it approaches its flagship GPT-6 Astra on several coding, computer-use, and professional-work evaluations at around one-fifth of Astra’s standard token prices. It is designed for tasks such as debugging code, working with complex documents, operating computer applications, and completing multi-step business processes. Lower costs matter when developers run agents repeatedly, because every step can consume tokens and increase the overall bill. Sol gives developers a way to build capable systems without always paying for the most expensive model.

Gemini 4 Argon: Built for Complex Workflows

Google DeepMind’s Gemini 4 Argon targets demanding tasks across software engineering, enterprise work, and cybersecurity. Google highlights its ability to handle long, multi-step tasks, understand different types of information, and find, validate, and patch software vulnerabilities. Its focus is not just producing an answer, but working through complicated problems that require several connected steps. Google also describes safeguards intended to reduce risks from malicious instructions and unsafe actions.

How Do They Compare?

The important difference is that these are not three directly equivalent products. Dots is an agent experience built around a model, while GPT-6.1 Sol and Gemini 4 Argon are AI models that can power applications and workflows. Sol emphasizes the balance between capability and cost, whereas Argon emphasizes complex reasoning, multimodal understanding, and defensive cybersecurity. Their published evaluations also use different tests, so the results should not be treated as a direct ranking.

What This Means for Developers

For developers, the next challenge is choosing the right combination of model, tools, permissions, and infrastructure. A lower-cost model can make frequent agent tasks more practical. A model built for complex workflows can help with demanding engineering problems. An always-on agent can connect these capabilities to everyday work and continue making progress without constant instructions.

But greater autonomy also creates new responsibilities. Agents need limited permissions, clear approval rules, monitoring, and human review for important actions. The direction is clear: AI is expanding beyond answering prompts toward completing real work, and the next stage will depend on making that work useful, affordable, and safe.

AI Agents leak 13000 screenshots - AI coding agents accidentally exposed more than 13,000 internal screenshots from over 300 organizations by creating public GitHub repositories to share images for code reviews. The screenshots included billing records, internal dashboards, and unreleased product features, with most stored under developers’ personal accounts. Read more.

DigitalOcean Builds a Home for Agents - DigitalOcean has launched Managed Agents in public preview, giving AI agents isolated microVM runtimes, persistent sessions, and managed infrastructure for running long-running tasks. Its Action Gateway also gives agents governed access to more than 16,000 tools through a single MCP endpoint, with credentials kept outside the agent. Read more.

Buzz of the Week!

Hazard Pointers

Hazard pointers are a lock-free memory reclamation technique used when concurrent threads access objects that may be deleted while another thread is still reading them. A thread publishes the address it is about to dereference in a hazard-pointer slot, telling other threads that the object must not be freed yet. The deleting thread first removes the object from the data structure, then checks whether any hazard pointer still references it. If one does, reclamation is postponed and the object is placed on a retire list for later cleanup. This avoids use-after-free bugs without requiring a global lock around every memory access. They are especially useful in lock-free queues, stacks, hash tables, and other high-performance concurrent data structures.

Things that launched. Things that went viral. Things you'll pretend to try.

benthos

benthos is a stream-processing tool for connecting APIs, databases, queues, and event streams without writing a full application.

meltano

meltano is open-source data integration platform for building and running ELT pipelines.

Dolt

Dolt is a SQL database with Git-like version control. You can branch, diff, merge, and revert database changes.

Build Braincells, Not Just Features

This weekend’s read: Introducing Kev

This week’s watch: Why Evolution made Men Bald??

Meanwhile…