7 min read · 14 sources

OpenAI just handed every developer the harness that powers Codex. The Agents API brings the same infrastructure that keeps OpenAI’s own coding agents alive for days to anyone willing to wire it into their stack. That’s not a wrapper around a chat model - it’s the orchestration layer, the environment management, the checkpointing that makes long-running agents practical. If you’ve been duct-taping together LangChain and cron jobs, this is the thing that replaces them.

The API provides infrastructure to keep agents running reliably for days, plus environments where agents can work with files, run code, and save intermediate results. Think of it as a managed runtime for agentic workloads: your agent gets a persistent workspace, can execute arbitrary code, and survives long enough to actually finish a task. For teams that have hit the wall where agents die after a few minutes or lose state between steps, this removes the hardest operational problem.

A software factory run with 7 tasks produced 90,000 lines of code changes over 15 hours with only 2 human actions - one to start, one to review.

The Agents API: Codex's Secret Sauce Goes Public

OpenAI’s Codex has been the company’s internal proving ground for agentic coding. The Agents API exposes that same harness - the loop that lets an agent plan, execute, observe, and iterate without human intervention. The infrastructure handles the boring parts that kill most agent projects: state persistence, environment isolation, and the ability to run for days without the whole thing collapsing into a zombie process.

For engineers, the interesting bit is what this changes operationally. Running agents at scale means managing compute, storage, and failure recovery. OpenAI is now selling that as a service. You bring the task, they bring the harness. The question is whether the API’s pricing and reliability hold up under production load, but the architectural pattern is clear: agents as a managed service, not a DIY project.

Anthropic Cuts Off AI Weapons Queries

Anthropic detected use of its models for AI weapon development and restricted access. The company says it still can’t determine whether a query serves a legitimate or nefarious purpose, so it erred on the side of caution - the consequences of mistakes could be severe.

This is the safety-vs-usefulness tension playing out in real time. Anthropic’s models are used by defense contractors, security researchers, and increasingly by people who want to build things that hurt others. The company can’t tell the difference reliably, so it’s restricting the whole category. Expect this to be an ongoing saga: every restriction pushes legitimate users toward other models, and every leak pushes regulators toward mandating restrictions.

Unitree: The First Humanoid Robot Stock

Source: chinatalk.media ↗

Unitree listed on the STAR Market on August 19, becoming the first humanoid robot stock on China’s A-shares market. Its market cap briefly reached 440 billion yuan - roughly $62 billion at current rates. That’s a staggering valuation for a company whose robots are still mostly demos and niche deployments.

The deep dive covers Unitree’s managerial structure and how the company is changing as it transitions from private startup to public company. The interesting engineering angle is cost control: Unitree has pushed robot prices down aggressively, and the public markets will now force quarterly discipline on a company that’s used to burning cash for R&D. If humanoid robots are going to be real products, this is the company that has to prove the economics work.

Will AI Push Growth to Double Digits?

Source: aleximas.substack.com ↗

Economists and analysts are increasingly projecting that AI could push economic growth to 10% to 15% per year. That’s not a typo - that’s the range being floated in serious analysis. For context, the US economy has been growing at roughly 2% annually. A jump to double digits would be the fastest economic expansion in recorded history.

For engineers, these projections imply rapid technological change that could reshape demand for their skills. If growth hits those levels, companies will be scrambling for talent, infrastructure, and new products. But treat the numbers with skepticism: the track record of economic forecasting is poor, and AI’s productivity gains have been slower to materialize than the hype suggests. The realistic take: AI will drive growth, but whether it’s 2% or 15% is anyone’s guess.

Tech Companies Move to Open AI Models

Source: blog.pragmaticengineer.com ↗

The Pragmatic Engineer covers how tech companies are moving to open AI models, with a focus on optimizing model selection and running benchmarks. Companies now benchmark all available frontier and open models weekly to build an accurate picture of their capabilities. They’re also caching prompts to save money when using Uber’s own harness, Minions.

The trend is clear: open models have closed the gap enough that companies can deploy them for production workloads. The engineering work is shifting from “which model is best” to “how do we route queries, cache aggressively, and evaluate weekly as new models drop.” More companies are setting per-developer limits on AI usage - the era of unlimited AI coding is ending as costs become real.

The Lifecycle of a Sharded Postgres Query

Source: planetscale.com ↗

PlanetScale’s post walks through what happens when a query hits a sharded database. A real database has hundreds of tables - the example keeps the schema simple with two tables spread across four shards. The post covers routing, distributed transactions, and the consistency tradeoffs that come with horizontal scaling.

The key insight for anyone running Postgres at scale: sharding changes everything about how you write queries. Joins become distributed operations, transactions need coordination, and the database’s behavior under load is fundamentally different from a single node. If you’re planning to shard, read this first.

The Software Factory Reality Check

Source: seangoedecke.com ↗

Two posts this week paint very different pictures of AI-driven development. One experiment spent 35 hours running a “software factory” that delivered “absolutely nothing of value.” Another run produced 90k LOC changes across 7 tasks with agents working 15 hours and only 2 human actions - one to start, one to review before deploy.

The difference isn’t luck. It’s harness design. The successful run used loops, graphs, and harnesses - structured approaches to getting quality out of agents. The failed run treated the factory as a black box. Armin Ronacher’s post on Astra gets at the same point: if you haven’t been thoroughly disappointed in AI capabilities, you haven’t tried enough. The “AI psychopath” problem is real - agents will do what you ask, including the wrong thing, with terrifying competence.

What Comes After Git

Source: ersc.io ↗

East River Source Control has been around for over a year but hasn’t been publicly discussed much. The post argues that source code is one of the most precious commodities an organization has, and that the rise of agentic development places particular strain on version control systems.

The argument: git was designed for human collaboration, not for thousands of AI agents making changes simultaneously. Agentic development needs version control that can handle massive parallel change sets, automatic conflict resolution, and review processes that scale beyond human capacity. Git’s model of branching and merging breaks down when the branch count is in the thousands.

They Really Do Think AI Might Kill Everyone

Source: seangoedecke.com ↗

The people building AI earnestly believe that it could kill us all. This isn’t a fringe position - it’s the stated belief of leadership at the major labs. The post examines what that means when the people building the technology also believe it’s an existential threat.

The engineering takeaway: the people who understand AI best are the most worried. That should inform how you deploy AI systems, what you let them do, and what guardrails you put in place. The “it’s just a tool” framing doesn’t hold when the tool’s creators think it might end civilization.

AI Deidentification Works - Maybe Too Well

Source: distributedthoughts.org ↗

AI can deidentify you better than many adversarial bot programs. The research shows that modern deidentification techniques can strip personal information from data more effectively than traditional methods - but that’s a double-edged sword. If AI can deidentify data perfectly, it can also potentially reidentify it.

Google's DMA Data Sharing Deadline

Source: desfontain.es ↗

Google has six months to implement the technical measures required by the EU’s Digital Markets Act around data sharing. The post examines what the compliance deadline means for how Google shares search data with third parties. For anyone building on search data, this could open new access - or it could result in Google finding creative ways to comply without actually sharing anything useful.

Seven Deadly Signs of AI Security Snake Oil

Source: grayswan.ai ↗

Gray Swan’s field guide covers the red flags every developer should catch before buying AI security tools. Your vendor says “OWASP LLM Top 10 covered” - that’s a checklist, not a security posture. The post identifies seven specific signs that an AI security product is selling hype instead of protection. Worth reading before your next vendor procurement cycle.

Loops, Graphs, and Harnesses: Getting Quality Out of a Software Factory

Source: ivokund.com ↗

The companion piece to the software factory discussion covers design principles for AI-driven development. Loops let agents iterate, graphs let them explore multiple paths, and harnesses constrain them to productive work. Most changes that would have made agents better at codebases make them worse - the post explains which design choices actually move the needle.