On Friday, OpenAI engineer Michael Bolin published a detailed technical breakdown of how the company's Codex CLI coding agent works internally, offering developers insight into AI coding tools that can write code, run tests, and fix bugs with human supervision. It complements our article in December on how AI agents work by filling in technical details on how OpenAI implements its "agentic loop."

AI coding agents are having something of a "ChatGPT moment," where Claude Code with Opus 4.5 and Codex with GPT-5.2 have reached a new level of usefulness for rapidly coding up prototypes, interfaces, and churning out boilerplate code. The timing of OpenAI's post details the design philosophy behind Codex just as AI agents are becoming more practical tools for everyday work.

These tools aren't perfect and remain controversial for some software developers. While OpenAI has previously told Ars Technica that it uses Codex as a coding tool to help develop the Codex product itself, we also discovered, through hands-on experience, that these tools can be astonishingly fast at simple tasks but remain brittle beyond their training data and require human oversight for production work. The rough framework of a project tends to come fast and feels magical, but filling in the details involves tedious debugging and workarounds for limitations the agent cannot overcome on its own.

Bolin's post doesn't shy away from these engineering challenges. He discusses the inefficiency of quadratic prompt growth, performance issues caused by cache misses, and bugs the team discovered (like MCP tools being enumerated inconsistently) that they had to fix.

The level of technical detail is somewhat unusual for OpenAI, which has not published similar breakdowns of how other products like ChatGPT work internally, for example (there's a lot going on under that hood we'd like to know). But we've already seen how OpenAI treats Codex differently during our interview with them in December, noting that programming tasks seem ideally suited for large language models.

It's worth noting that both OpenAI and Anthropic open-source their coding CLI clients on GitHub, allowing developers to examine the implementation directly, whereas they don't do the same for ChatGPT or the Claude web interface.

An official look inside the loop

Bolin's post focuses on what he calls "the agent loop," which is the core logic that orchestrates interactions between the user, the AI model, and the software tools the model invokes to perform coding work.

As we wrote in December, at the center of every AI agent is a repeating cycle. The agent takes input from the user and prepares a textual prompt for the model. The model then generates a response, which either produces a final answer for the user or requests a tool call (such as running a shell command or reading a file). If the model requests a tool call, the agent executes it, appends the output to the original prompt, and queries the model again. This process repeats until the model stops requesting tools and instead produces an assistant message for the user.

That looping process has to start somewhere, and Bolin's post reveals how Codex constructs the initial prompt sent to OpenAI's Responses API, which handles model inference. The prompt is built from several components, each with an assigned role that determines its priority: system, developer, user, or assistant.

The instructions field comes from either a user-specified configuration file or base instructions bundled with the CLI. The tools field defines what functions the model can call, including shell commands, planning tools, web search capabilities, and any custom tools provided through Model Context Protocol (MCP) servers. The input field contains a series of items that describe the sandbox permissions, optional developer instructions, environment context like the current working directory, and finally the user's actual message.

As conversations continue, each new turn includes the complete history of previous messages and tool calls. This means the prompt grows with every interaction, which has performance implications. According to the post, because Codex does not use an optional "previous_response_id" parameter that would allow the API to reference stored conversation state, every request is fully stateless (that is, it sends the entire conversation history with each API call rather than the server retrieving it from memory). Bolin says this design choice simplifies things for API providers and makes it easier to support customers who opt into "Zero Data Retention," where OpenAI does not store user data.

The quadratic growth of prompts over a conversation is inefficient, but Bolin explains that prompt caching mitigates this issue somewhat. Cache hits only work for exact prefix matches within a prompt, which means Codex must carefully avoid operations that could cause cache misses. Changing the available tools, switching models, or modifying the sandbox configuration mid-conversation can all invalidate the cache and hurt performance.

The ever-growing prompt length is directly related to the context window, which limits how much text the AI model can process in a single inference call. Bolin writes that Codex automatically compacts conversations when token counts exceed a threshold, just as Claude Code does. Earlier versions of Codex required manual compaction via a slash command, but the current system uses a specialized API endpoint that compresses context while preserving summarized portions of the model's "understanding" of what happened through an encrypted content item.

Bolin says that future posts in his series will cover the CLI's architecture, tool implementation details, and Codex's sandboxing model.

Read full article

Comments

Yann LeCun helped create the technology behind today’s chatbots. Now he says many tech companies are on the wrong path to creating intelligent machines.

Concerns about the affordability of education, housing, health care, having a family and retirement are driving economic anxieties, a New York Times/Siena poll found.

Every spring, raptors return to nesting sites across northern Michigan. The smallest of these birds of prey, a falcon called the American kestrel (Falco sparverius), flies through the region’s many cherry orchards and spends its days hunting for even tinier creatures to eat. This quest keeps the kestrels fed, but it also benefits the region’s cherry farmers.

Fruit farmers have been working symbiotically with kestrels for decades, adding nesting boxes and reaping the benefits of the birds eliminating the mice, voles, songbirds, and other pests that wreak havoc by feeding on not-yet-harvested crops. In addition to limiting the crop damage caused by hungry critters, new research suggests kestrels also lower the risk of food-borne illnesses.

The study, published in November in the Journal of Applied Ecology, suggests the kestrels help keep harmful pathogens off of fruit headed to consumers by eating and scaring off small birds that carry those pathogens. Orchards housing the birds in nest boxes saw fewer cherry-eating birds than orchards without kestrels on site. This translated to an 81 percent reduction in crop damage—such as bite marks or missing fruit—and a 66 percent decrease in branches contaminated with bird feces.

“Kestrels are not very expensive to bring into orchards, but they work pretty well” at deterring unwanted bird species, said Olivia Smith, lead study author and assistant professor of horticulture at Michigan State University. “And people just like kestrels a lot, so I think it’s an attractive strategy.”

Finding a good strategy for managing pests is essential for cherry farmers. Pests cause expensive damage that worsens yields already impacted by other threats to the cherry industry, such as climate change, labor shortages, and the vagaries of international trade. To stop the added damage from pests, growers have turned to nets covering their trees, noisemakers, scarecrows, pesticides, and even the removal of natural habitats around crop-growing areas.

However, these options can be expensive and aren’t always effective. Even with these management strategies in place, birds like starlings, robins, and crows cost farms in some top cherry growing states—including Michigan, New York, Oregon, Washington, and California—about $85 million annually. For many growers, this is where the kestrels come in.

It may seem counterintuitive to solve a bird problem by bringing in more birds, but kestrels are skilled hunters whose presence drives off songbirds afraid of being eaten. Habitat loss, competition for food, and climate change are leading to slow and steady population declines for the American kestrel, losses of about 1.4 percent annually. Still, these birds are abundant enough that, in many areas of the continental United States, all farmers need to do to attract them is add a nesting box to their land.

“I’ve noticed a difference having the kestrels around, hovering over the spring crops,” said Brad Thatcher, a farmer based in Washington state who has housed kestrels on April Joy Farm, an organic fruit and vegetable farm, for over 13 years. “There’s very little fecal damage from small songbirds at that time of year versus the fall.”

With farmers who already had kestrels on their land reporting fewer songbirds and less crop damage, study authors hypothesized that food safety risks associated with pathogens birds carry may also be lower for farms harboring kestrels. To test this, the researchers evaluated 16 sweet cherry orchards in Michigan’s Leelanau and Grand Traverse counties (the latter of which is considered the “Cherry Capital of the World”). Eight of the orchards studied had nesting boxes for kestrels and eight did not.

The study authors randomly selected two areas of each orchard to search for crop damage and fecal contamination. The orchards frequented by kestrels saw the amount of damaged fruit drop from 2.5 percent to 0.47 percent. The number of crops contaminated by bird droppings also saw a three-fold decrease, falling from 6.88 percent to 2.33 percent. When researchers tested this excrement, they found that more than 10 percent contained campylobacter, a type of bacteria that is common in birds and causes food-borne illness in humans.

Campylobacter is a common cause of food poisoning and is on the rise in Michigan and around the world. It spreads to humans through food products made from, or that come into contact with, infected animals, primarily chickens and other birds. So far, only one outbreak of campylobacteriosis has been definitively linked to feces from wild birds. Still, because it causes milder symptoms than some other types of bacteria, the Centers for Disease Control considers campylobacter a significantly underreported cause of food-borne illness that may be more common than current data indicates.

“Trying to get more birds of prey would be beneficial to farmers,” Smith said. “If you have one predator, versus a bunch of prey, you have fewer birds overall. If you have a lot fewer birds, even if the ones that are there are carrying bacteria, then you can reduce the transmission risk.”

The study’s findings that kestrels significantly reduce physical damage and food safety risks on Michigan cherry farms demonstrate that managing crops and meeting conservation goals—by bolstering local kestrel populations and eliminating the need to clear wildlife habitat around agricultural areas—can be done in tandem, study authors say. They recommend farmers facing pest-management issues consider building kestrel boxes, which cost about $100 per box and require minimal maintenance.

Whether nesting boxes in a given region will be successfully inhabited by kestrels depends on whether there is an abundance of the birds there. In Michigan’s cherry-growing region, kestrels are so abundant that 80 percent to 100 percent of boxes become home for kestrels rather than other nesting birds, said Catherine Lindell, avian ecologist at Michigan State University and senior author of the study.

“It seems like this is just a great tool for farmers,” Lindell said, suggesting interested farmers “put up a couple boxes and see what happens.”

K.R. Callaway is a reporter and editor specializing in science, health, history, and policy stories. She is currently pursuing a master’s degree in journalism at New York University, where she is part of the Science, Health, and Environmental Reporting Program (SHERP). Her writing has appeared in Scientific American, Sky & Telescope, Fast Company and Audubon Magazine, among others.

This story originally appeared on Inside Climate News.

Read full article

Comments

New court documents reveal even more ways DOGE improperly accessed and shared sensitive personal data last year — and how that data appears to have been used to advance dubious fraud claims.

(Image credit: Kayla Bartkowski)

Earlier this week, I found myself with time to kill in an airport lounge ahead of my flight back to San Francisco from the World Economic Forum in Davos, Switzerland.

So I did what any bleary-eyed tech editor who felt totally wiped out from dozens of interviews and events would do: I parked myself in an oversize lounge chair, booted up the airport Wi-Fi and decided to see if Claude Cowork, Anthropic’s new AI productivity tool, could write a presentation I planned to give, summarizing the trip.

Specifically, I wanted to see how the AI would analyze transcripts from most of my conversations, pull out the key themes and put them in a slide deck for a presentation I intended to give Pro subscribers to The Information in the coming days.

I learned a ton of lessons.